The

Technology
of
Computer
Music
MAX V. MATHEWS
with the
collaboration
of
Joan E. Miller
F. R. Moore
J. R. Pierce
and J. C. Risset
The
Technology
of
Computer
Music

THE M.I.T. PRESS

Massachusetts Institute of Technology
Cambridge, Massachusetts, and London, England
 Copyright © 1969 by
The Massachusetts Institute of Technology

,All rights reserved. No part of this book may be

reproduced or utilized in a,ny form or by any means,
electronic or mechanical, including photocopying,
recording, or by any information storage and retrieval
system, without permission in writing from the publisher.

Second printing, January 1974

ISBN 0262 130505

Library of Congress catalog card number: 69-12754

Printed and bound in the United States of America

Contents

1. Fundamentals
Introduction 1
Numerical Representation of Functions of Time 2
Sampling and Quantizing 4
Foldover Errors 7
*Mathematical Analysis of Sampling 11
*Alternative Analysis of Sampling 16
Bounding Sampling Errors 18
*Sample and Hold Analysis 21
*Analysis of Quantizing Errors 22
Digital-to-Analog and Analog-to-Digital Converters 25
Smoothing-Filter Design 29
Digital Data Storage and Retrieval for Sound 31
Fundamental Programming Problems for Sound Synthesis 33
Overview of Sound-Synthesis Program-Music V 37
Annotated References by Subject 39
Problems for Chapter 1 40

2. A Sequence of Tutorial Examples of Sound Generation

Introduction 43
The Simplest Orchestra 44

v
vi CONTENTS

Simple Unit Generators to Output, Add, and Multiply 46

Examples of Simple Instruments 53
C0NVT Function to Process Note Parameters 62
Additional Unit Generators-RAN, ENV, FLT 68
Composing Subroutines-PLF 78
Compositional Functions 86
Pass II Subroutines-PLS Pitch-Quantizing Example 94
Interactions Between Instruments 98
Parting Exhortations to the Student 104
Annotated References by Subject 105
Problems for Chapter 2 105

3. Music V Manual
1. Introduction 115
2. Description of Pass I 120
3. Operation Codes and Corresponding Data Statements 122
4. Definition of Instruments 124
5. Unit Generators 126
6. Special Discussion of 0SC Unit Generator 134
7. Input-Output Routines for Pass I and Pass II 139
8. PLF Subroutines 144
9. General Error Subroutine 144
10. Description of Pass II 145
11. WRITE2 148
12. C0N-Function Evaluator for Line-Segment Functions 150
13. S0RT and S0RTFL 150
14. PLS Routines 151
15. C0NVT-Convert Subroutine 152
16. Description of Pass III 153
17. I and IP Data Arrays in Pass III 158
18. Note Parameters 160
19. Instrument Definition 161
20. F0RSAM 162
21. SAMGEN 165
22. SAM0UT 167
23. SAM0UT for Debugging 167
24. Acoustic-Sample Output Program: FR0UT 167
25. GEN-Pass III Function-Generating Subroutines 169
 CONTENTS vii

26. Data Input for Pass III-DATA 172

Annotated References by Subject 172

Appendix A. Psychoacoustics and Music 173

Appendix B. Mathematics 180
Index 184
1 Fundamentals

Introduction
This book is intended for people who plan to use computers for
sound processing. Present users range from engineers and physicists
concerned with speech and acoustics to musicians and phoneticians
concerned with sound synthesis and speech production and perception.
The widely varied technical and mathematical background of this
audience makes it hard to select a technical level for this presentation.
Some experience with a computer language such as F0RTRAN will be
assumed, though it could be obtained at the time this material is studied.
Occasionally a satisfactory explanation of some point requires
mathematics at the level of a graduate curriculum in electrical engineer-
ing. These mathematical sections have been quarantined and marked
with an asterisk. Although the mathematical material adds essential
understanding of sound processing, the rest of the book is intended to
be comprehensible without it. The implications of the mathematics are
usually given in elementary terms in other sections. Also, Appendix B
lists the main relationships required for mathematical background.
Chapter I covers some fundamentals that are basic to all computer
sound processing-the representation of sounds as numbers, the under-
lying processes of sampling and quantizing a sound wave, the approxi-
mations and errors that are inherent in sampling and quantizing, the
operation of digital-to-analog and analog-to-digital converters, the
2 CHAPTER ONE

construction of smoothing filters, the storage and manipulation of

sound waves in numerical form, and, last, an introductory look at the
computer programming for sound processing which is the central
subject of the rest of the book.

Numerical Representation of Functions of Time

Sound can be considered as a changing or time-varying pressure in the
air. Its subjective characteristics, how it" sounds," depend on the specific
way the pressure varies. For example, a tone with a definite pitch A
above middle C has a periodic pressure variation that repeats itself 440
times each second. A constant pressure is heard as silence.
Since the essence of the sound depends on the nature of the variations
in pressure, we will describe a sound wave by a pressure function pet).
The p stands for pressure, the t for time, and the parentheses indicate
that pressure is a function of time; in other words, pressure changes as
time goes on. The term pet) will represent the pressure function of time
or, more briefly, the pressure function.
One way to describe pressure functions is to draw a picture or graph
showing how they vary with time. Two simple examples are shown in
Fig. 1. Figure la shows a constant pressure heard as silence. Figure Ib
shows a sinusoidal variation that repeats itself each 1/500 second. Thus
the pitch will be slightly below C (524 Hz). The time scale is labeled in
thousandths of a second. The variations are very rapid compared with
the times in which we schedule our lives. The pressure wave cannot be
described as a single quantity or number. Its whole history must be
drawn, and there is an infinite variety of ways in which it can change
from millisecond to millisecond.
The pressure is shown increasing or decreasing around zero pressure.
Actually the variations are around the pressure of the atmosphere, about
15 pounds per square inch. However, atmospheric pressure is essentially
constant and produces no sound. The variations are small compared to
the atmospheric pressure: A very loud sound would change from 15 to
15.001 pounds per square inch. The minuteness of this variation
indicates the great sensitivity of our ears.
All sounds have a pressure function and any sound can be produced
by generating its pressure function. Thus if we can develop a pressure
source capable of producing any pressure function, it will be capable of
producing any sound, including speech, music, and noise. A digital
computer, plus a program, plus a digital-to-analog converter, plus a
loudspeaker come close to meeting this capability.
 FUNDAMENTALS 3

Jt----,-,-,-,-,-,-"., .
.001 .002 .003 .004 .005 .OOG .007
Time in seconds

(0)

Fig. 1. Simple pressure

functions: (a) silence; (b)
500-Hz sinusoid.

Time in seconds

(b)

In the past most sounds have originated from the vibrations and
movements of natural objects-human vocal cords, violin strings,
colliding automobiles. The nature of these sounds is determined by and
limited by the particular objects. However, in the last 50 years the
loudspeaker has been developed as a general sound source. It produces
a pressure function by means of the vibrations of a paper cone actuated
by a coil of wire in a magnetic field. The movement of the cone as a
function of time, and hence the resulting pressure function, are deter-
mined by the electric voltage (as a function of time) applied to the coil.
Loudspeakers are not perfect: they distort all sounds slightly, and some
sounds are hard to produce. However, the almost universal range of
sounds they generate in a satisfactory way is demonstrated by the range
of sounds that can be played on phonograph records and on radios.
Loudspeakers are sound sources of almost unlimited richness and
potential.
To drive a loudspeaker and produce a desired pressure function, an
electric voltage function of time must be applied to its coil. Exchanging
the problem of generating a pressure function for generating a voltage
function might seem to offer little gain. However, very versatile methods
exist for producing electric functions.
4 CHAPTER ONE

One popular method of generating a great variety of voltage functions

is the phonograph record. The minute wiggles in the grooves on the
record are converted into a motion function of the needle on the pickup.
The wiggles are a space function, but this space function is convert¢ to
a time function by the turntable, which moves the groove past the
needle at a particular speed. The motion function of the needle is
converted to a voltage function in one of a number of well-known ways.
The voltage, after amplification, is applied to the loudspeaker.
The value of the phonograph as a source of voltage functions is that
a wiggle of almost any shape can be cut in the groove. If one had a
minute chisel, grooves for new sounds could be cut by hand. However,
the computer can accomplish an equivalent result by a much easier
process.

Sampling and Quantizing

The pressure functions that we hear as sound are generated by
applying the corresponding voltage functions to a loudspeaker. How
can voltage functions be produced from the numbers in a computer?
The process is shown in Fig. 2. Numbers stored in the computer
memory are successively transferred to a digital-to-analog converter.

Loudspeaker
Computer
I---Me~ory ---~ Digital- Smoothing
to-
: 6,13,16,12,11,15,: filter
analog
L13~~:~''':''':'' __ J converter
o to 5KHz

Sequence of Sequence of Sound pressure

num bers from pulses with wOve obtained
computer amplitude by smoothing
memory proportional pulses
to numbers

.
:::J
20
+
613.161211 15 .. ·
II>
......~
II>

.8
II>

10
~2
~g
00
~t -10
o
a.
~-200~------------~1------------~2
Time (milisecondsl

Fig. 2. Computer-to-pressure conversion.

 FUNDAMENTALS 5

F or each number the converter generates a pulse of voltage whose

amplitude equals the number. These pulses are shown on the graph in
the lower part of Fig. 2. The square corners of the pulses are smoothed
with a filter (low-pass filter) to produce the smooth voltage function
drawn through the pulses in Fig. 2. This voltage, supplied to the
loudspeaker, produces the desired pressure wave.
We will discuss later the electrical details of the digital-to-analog
converter and the smoothing filter, and analyze the errors quantita-
tively. Here we would like to give some physical intuition about the
process. From inspection of Fig. 2, it seems evident that a great variety
of time functions can be generated from the right numbers. Each number
simply gives the value of the function at one instant in time. Practically,
the computer can produce any set of numbers and hence any time
function. However, some functions are more difficult to produce than
others, and certain approximations are involved in producing any
function. It is important to understand the nature of these approxima-
tions in order to use the computer as an effective sound source. Sampling
and quantizing are the two approximations involved in representing a
continuous function by a set of numbers.
A continuous function of time can change at every instant in time.
The numbers, by contrast, are ·converted to pulses which are constant
for a given duration, called the pulse width or sampling time. In Fig. 2,
there are 10 pulses each millisecond, so the sampling time is 1/10,000
sec. It is often convenient to talk about the sampling rate that is
l/(sampling time). Thus a sampling time of 1/10,000 sec corresponds to
a rate of 10,000 samples per second.
Intuitively it seems that, if we make the sampling time very small, the
pulses will be a good approximation to the continuous function as
illustrated in Fig. 3a, and if we make the sampling time large, as in
Fig. 3b, we will get a poor approximation. Of course, the approximation
depends on the function too. More pulses are needed to approximate a
rapidly changing function than a slowly changing one. The rapidly
changing function is best thought of as having higher frequencies than
the slowly changing function. Thus a higher sampling rate, and hence
more pulses, and hence more numbers, will be required to approximate
high-bandwidth (hi fi) sound than low-fidelity sound.
Mathematically it has been shown that R pulses per second are
needed to approximate perfectly a function with a bandwidth R/2
cycles per second. Thus, to approximate a high-fidelity sound with a
bandwidth of 15,000 Hz, we require 30,000 samples per second, or a
sampling time of 1/30,000 sec.
6 CHAPTER ONE

Pressure

Time

Fig. 3. Example of
various sampling rates:
(0 )
(a) high sampling
rate ; (b) low sampling
Pressure rate.

Time

(b)

We can now begin to appreciate the huge task facing the computer.
For each second of high-fidelity sound, it must supply 30,000 numbers
to the digital-to-analog converter. Indeed, it must put out numbers
steadily at a rate of 30,000 per second. Modern computers are capable
of this performance, but only if they are expertly used. We can also
begin to appreciate the inherent complexity of pressure functions
producing sound. We said such a pressure could not be described by
one number; now it is clear that a few minutes of sound require
millions of numbers.
The second approximation is called quantizing. The numbers in
computers contain only a certain number of digits. The numbers in the
Fig. 2 computer have only two digits. Thus, for example, all the pulse
amplitudes between 12.5 and 13.5 must be represented by the number
13. Of course we could build a larger computer that could handle
three-digit numbers. This machine could represent 12.5 exactly. How-
ever, it would have to approximate all the amplitudes between 12.45
and 12.55 by 12.5. Furthermore, the more digits, the more expensive
will be the computer.
 FUNDAMENTALS 7

The quantizing errors are closely equivalent to the noise and distor-
tion that are produced by phonographs, tape recorders, amplifiers, and
indeed all sound-generating equipment. Their magnitude can be
estimated in terms of signal-to~noise ratios or percentage distortions.
The approximate signal-to-noise ratio inherent in a given number of
digits equals
Maximum number expressible with the digits
Maximum error in representing any number
For example, with two-decimal digits, the maximum number is 99 and
the maximum error is .5. The signal-to-noise ratio is

~~ ~ 200 or 46 dB

Three-decimal digits would correspond to a signal-to-noise ratio of

999/.5, or 66 dB. This ratio is as good as very high-quality audio
equipment. Three-decimal digits would be a very small number of
digits for most computers; hence quantizing errors are not critical.
Sampling rate, by contrast, is often critical.
Most computers use binary rather than decimal numbers. The same
method can be used to estimate quantizing errors. For example,
10-digit binary numbers can express the decimal integers from 0 to
1023. Hence the equivalent signal-to-noise ratio is 1023/.5, or about
66 dB. Typically, 10- to 12-binary-digit numbers are used for sound
generation.
The limit to the quantizing errors is usually the digital-to-analog
converter rather than the computer. Computers with an accuracy of
12 to 36 digits or more are standard. Converters with accuracy beyond
12 digits are difficult to make. Twelve digits would correspond to a
signal-to-noise ratio of 78 dB. Although this ratio may seem more than
sufficient, its quantizing noise is occasionally objectionable when very
soft sounds are generated, or when a peculiar interaction arises between
sounds and noise. Thus it seems prudent to use at least 12 digits.

Foldover Errors
The generation of voltage functions from quantized samples is a
practical, powerful, and useful method when coupled to modern com-
puters. Most of this book is concerned with applications of this method.
In order to use the method, its errors and limitations must be under-
stood and avoided. A mathematical analysis of the errors is given later
8 CHAPTER ONE

in this chapter. Since the quantizing noise is similar to noise in other

apparatus, it usually causes no unexpected trouble. However, sampling
produces a frequency distortion called fo/dover, which can generate
surprising and unwanted frequencies. Because of foldover's insidious
nature, we will present an intuitive discussion here, in addition to an
analysis later.
The limitations of sampling in reproducing waveforms that contain
very high frequencies can be illustrated graphically. Figure 4a shows the
sampling of a periodic sequence of short pulses. We see that only one
out of the four pulses shown overlaps a sampling time, and hence in the
figure all samples but one are zero in amplitude. The sampling of the
regular sequence of pulses produces samples spaced much farther apart
in time than the pulses are.
Figure 4b is another illustration of the defects of sampling. Here
sampling of a square wave produces pairs of positive samples separated
by single negative samples.
Thus sampling a waveform can produce samples that represent the

Voltage function

...Jl n'--_--'n. . .__fL.....

Sampling times

I I I
Sample pulses

~--------------------~.Time

(a)

Fig. 4. (a) Sampling a

function that has
narrow pulses; (b)
Voltage function sampling a square-
wave function.
Hllrill
LJ LJ LJ L-
Sampling times

I I I
Somple pulses

Time

(b)
 FUNDAMENTALS 9

waveform poorly. This happens when the voltage function has fre-
quencies higher than Rj2 Hz, where R is the sampling rate. This is the
case for the voltage functions and sampling rates shown in Fig. 4.
When the voltage function contains frequencies higher than Rj2 Hz,
these higher frequencies are reduced, and the resulting sound is heard
somewhere in the range 0 to Rj2 Hz. For example, if the sampling rate
is 30,000 Hz and we generate samples of a sine wave at a frequency of
25,000 Hz
sin (27T·25,000·t)
the resulting voltage function out of the low-pass filter (smoothing
filter, Fig. 2) will be a sine wave at 5000 Hz
sin (27T' 5000· t)
More generally, if we generate samples of a sine wave at F Hz, where F
is greater than Rj2, the resulting frequency will be
FroId = R - F
The frequency F is reflected or folded by the sampling frequency; hence
the term foldover.
Why does fold over occur? Some physical feeling is suggested by
Fig. 5. Here we have diagrammed the example discussed above, of a
25,000-Hz sine wave sampled at 30,000 Hz. The samples of the 25,000-
Hz wave are shown as points, and the actual numbers are
1, .5, - .5, -1, - .5, .5, 1, .5, ...

samPling time-33~fLsec

::;1

GI
sin (217"'25.000 t)
rT (Sampling rate = 30.000Hz)

~0~~-+-+'~~~~~+-~-4~~~~~~
GI Time in
0. microseconds
-I

sin (217"·5000t )

Samples (I •. 5.-.5.-1.-.5 •. 5.1.···)

Fig. 5. Example of high-frequency (25,000 Hz) and foldover frequency

(5000 Hz) resulting from low sampling rate (30,000 Hz).
10 CHAPTER ONE

A 5000-Hz sine wave is also shown, and it also passes through the same
sample points. In other words, the 5000-Hz wave will have the identical
samples and therefore the identical numbers as the 25,000-Hz wave.
When the pulses produced by these numbers are put into the low-pass
filter, a 5000-Hz wave will come out, because the low-pass filter passes
low frequencies and attenuates high frequencies.
The essential point in the example is the identity of the samples of the
25,000-Hz and 5000-Hz waves. Hence from the samples there is no way
to distinguish between these frequencies. No computer program or
electric filter or other device can separate identical objects. For practical
purposes, the digital-to-analog converter and smoothing filter will
always be designed to interpret the samples as a 5000-Hz wave, that is, a
wave between 0 and Rj2 Hz. Thus one must be willing to accept this
frequency in the sound, or one must avoid generating samples of a
25,000-Hz wave (in general, a wave with frequencies greater than
Rj2 Hz).
The example chosen was simple in order that the graph could be
easily seen and the numbers easily computed. But the relation
F fOId = R - F holds for sine waves generally. More complex periodic
waves can be decomposed into individual harmonics and the foldover
frequency calculated separately for each harmonic.
Foldover also occurs from mUltiples of the sampling rate. Com-
ponents of ± R ± F, ± 2R ± F, ± 3R ± F, etc., are produced by the
digital-to-analog converter. However, in most cases only R - F is
troublesome.
We will next illustrate the sound of fold over with two examples.
Suppose a sine wave with continuously increasing frequency (glissando)
is sampled. What will be heard? As the frequency increases from 0 to
15,000 Hz, an increasing frequency going from 0 to 15,000 Hz will be
heard. But as the frequency increases from 15,000 to 30,000 Hz, a
decreasing frequency (30,000 - F) will be heard, going from 15,000 to
o Hz. This is usually a shock! If we persist in raising the frequency and
proceed from 30,000 to 45,000 Hz, the resulting sound will go upward
from 0 to 15,000 Hz (- 30,000 + F).
If we generate a complex tone with a high pitch and many harmonics,
the higher harmonics will fold over and appear at unwanted frequencies.
For example, the fifth harmonic of a 3000-Hz tone will occur at 15,000
Hz. That is the highest frequency that is not folded at a 30,000-Hz
sampling rate. The sixth harmonic (18,000 Hz) will be generated at
12,000 Hz and thus add to the fourth harmonic. The ninth harmonic
(27,000 Hz) will appear at the fundamental frequency, 3000 Hz.
 FUNDAMENTALS 11

In this example the fundamental frequency is a divisor of the sampling

rate. In this case the folded higher harmonics fall at exactly the fre-
quencies of the lower-frequency harmonics, thus producing a slight
distortion of the spectrum shape. Such distortion is seldom objection-
able. However, the sampling rate is not usually an exact multiple of the
fundamental frequency, and the folded harmonics will not fall on lower
harmonics. For a tone whose fundamental is 3100 Hz, the sixth har-
monic (18,600 Hz) will fall at 11,400 Hz between the third harmonic
(9300 Hz) and the fourth harmonic (12,400 Hz). At the least, the tone
quality will be much changed. At the worst, dissonance that resembles
intermodulation distortion will be generated.
The practical conclusion from this discussion of foldover is: avoid
generating samples of waveforms whose frequencies are higher than
half the sampling rate.

*Mathematical Analysis of Sampling

We will present a version of the sampling theorem in this section to
show that frequency-limited functions can be represented by a sequence
of numbers, and to show what errors are made by sampling functions
that are not strictly frequency limited. The main result is: functions
containing frequencies between 0 and R/2 can be exactly represented by
R samples per second. The sampling of functions that are not frequency
limited produces fold over errors whose magnitude can be calculated.
Also errors introduced by the smoothing filter can be calculated.
Quantizing errors will not be considered: each sample will be assumed
to be exactly represented by a number with infinite decimal places.
Figure 6 presents a block diagram of a sampling and desampling
process which we will use to analyze sampling. A time function

p (iT) i ="', -1,0,1",'

.------,
() Sampler zIt) T·z(f) Smoothing pll(t)
p t (aoolog- Impulse Amplifier
filter
digital modulator (gain T)
F(w)
converter)

Fig. 6. Conceptual block diagram of sampling-desampling process.

p(t), -00 < t < 00, is sampled. The analog-to-digital converter produces
a sequence of numbers p(iT), i = ... , -1,0, 1,2, ... , equal to p(t) at
the sampling times iT. The sampling interval is T, and the sampling
rate R = liT.
12 CHAPTER ONE

The desampling process consists of an "impulse" modulator, an

amplifier, and a smoothing filter. The output of the modulator is a
sequence of impulses z(t) whose areas are respectively proportional to
p(iT). Thus
+00

z(t) = L
i= - 00
8(t - iT) p(iT) (1)

where 8(t) is a unit impulse at time t = O. In actual practice, the impulse

modulator is well approximated by a pulse modulator producing
pulses of finite width. The impulses are smoothed by an ideal low-pass
filter, having unity gain from 0 to R/2 Hz and zero gain above R/2 Hz.
Such a filter is unrealizable but can be well approximated by filters
that can be constructed.
The output of the smoothing filter p*(t) should equal or closely
approximate pet). The essential result is simply that, for any frequency-
limited pet), pet) = p*(t). This result not only establishes that any
frequency-limited function can be generated from samples but also
gives a recipe for sampling any frequency-limited function. Thus a basis
is built for both sampling and desampling. Although one frequently
works with only sampling (for sound analysis) or only desampling
(for sound synthesis), the complete sampling-desampling process is
conceptually necessary to establish the validity and accuracy of either
process.
In addition to being frequency limited, pet) must satisfy certain
additional requirements of a more subtle mathematical nature. In
general we will ignore these qualifications here, with apologies to the
mathematicians. Functions obtained from sound waves satisfy these
qualifications.
The function pet) can be expressed as the inverse Fourier transform
1 f+OO P(w)e jwt dw
pet) = -27T (2)
_ 00
where the spectrum ofp(t) is pew). Ifp(t) is frequency limited to half the
sampling rate R, then pew) = 0 for Iwl ~ -1Wo where Wo = 27TR. A

P(w)

r:h o
w rod Isec
..
Fig. 7. Typical frequency-
limited spectrum .
 FUNDAMENTALS 13

sketch of such a pew) is shown in Fig. 7. The output of the impulse

modulator may be written as the product
z(t) = m(t)· pet)
where met) is a sequence of unit impulses at the sampling rate
+00

met} =
1=
L -00
8(t - iT)

as shown in Fig. 8. The spectrum M(w) of met) can be formally

m(tl

... t t
-2T -T
it t t
0 T
t sec
2T 3T
..
Fig. 8 .
pulses .
Sampling im-

represented 1 as a sequence of impulses in the frequency domain

M(w) =.;
2
n=
L
+00

-00
8(w - nwo) (3)

as shown in Fig. 9.

To
w rod/sec
t
W·o

Fig. 9. Spectrum of sampling impulses.

1 This spectrum may be formally derived from the Fourier series analysis of
met), which yields
1 2 ex>
met) = T + T n~l cos nWot
The spectrum of cos nWot is
1T[a(W - nwo) + a(w + nwo)]
Hence the spectrum of met) may be computed as the sum of spectrums of cos nWot
terms
14 CHAPTER ONE

Using the convolution theorem, the spectrum Z(w) of z(t) may be

written in terms of M(w) and P(w) as the integral
I f+oo
Z(w) = 27T _ 00 M(o:)P(w - 0:) do: (4)

Substituting the series for M(w)

f+oo
Z(w) =
1
Tn h+00

00 _ 00 B(0: - nwo)P(w - 0:) do:

which, because of the nature of the impulse function, simplifies to

1 + 00
Z(w) = T L
n= - 00
pew - nwo) (5)

The spectrum P*(w) of the output p*(t) is Z(w) times the product of the
amplification T and the transfer function F(w) of the smoothing filter
+00

P*(w) = F(w)
n= -
L 00
pew - nwo) (6)

Equation 6 is the basic result and holds for both frequency-limited and
frequency-nonlimited P(w),s. It says that P*(w) contains the sum of
pew) spectra which have been shifted by nwo. Let us examine P*(w)
for the frequency-limited case.
Figure 10 shows a sketch of T· Z(w) for the P(w) shown in Fig. 7.
°
Since P(w) = for Iw I ~ wo/2, the sum of shifted P(w) spectra gives

Ideal smoothing- filter

transfer function

w rad/sec

Fig. 10. Spectrum of T·Z(w) and smoothing-filter transfer function with

frequency-limited function.

copies of the pew) spectra centered at ... , -Wo, 0, we' 2wo, ... rad/sec.
If the smoothing transfer function F(w) is such that F(w) = 1 for
°
Iwl < wo/2, and F(w) = for Iwl ~ wo/2 as shown in Fig. 10, then
P*(w) is simply the center hump ofT·Z{w). Geometrically it is easy to
see that P*(w) = pew) and therefore that pet) = p*(t).
 FUNDAMENTALS 15

Under these same conditions, Eq. 6 reduces to

P*(w) = pew) (7)
The required conditions for Eq. 7 to hold are again
pew) = 0 for Iwl 2:: wo/2, and F(w) = 1 for Iwl < wo/2
F(w) = 0 for Iwl 2:: wo/2

Thus we have established our main claim and shown how a faithful
replication of any frequency-limited function can be generated from
samples.
What errors are produced if pew) is not frequency-limited? Figure 11
shows such a case. P(w) is nonzero until w equals .9wo. The summation

w rod/sec

Fig. 11. Spectrum of T·Z(w) with function having a too wide frequency
spectrum.

specified by Eq. 5 causes the tail (P(w), wo/2 < w < wo) to add energy
to Z(w) in the frequency region 0 < w < wo/2. The tail is said to be
folded around wo/2, and hence the distortion is called foldover. Energy
in P(w) at frequencies w appears in P*(w) at frequencies Wo - w. This
distortion is produced by the terms P(w - wo) and P(w + wo) in
Eq. 5. If pew) contains even higher frequencies, distortions with
frequency shifts of 2wo - w will be introduced by the P(w - 2wo) and
pew + 2wo) terms, and so forth.
In addition to fold over, errors are also introduced by the smoothing
filter. The transfer function F(w) is one term in Eq. 6. Realizable filters
cannot achieve the ideal transfer function of unity for Iwl < wo/2 and
zero for Iwl 2:: wo/2. A typical function is sketched in Fig. 11. Two
types of errors are caused. Departures of the amplitude from unity for
Iw I < wo/2 distort P*(w) within the band of interest and produce
in-band distortion. These distortions are typical of errors in other
electronic equipment and are often measured in decibels of departure
from unity or "flatness." Flatness within ± I dB is typical and easy to
produce.
16 CHAPTER ONE

Departures of the amplitude from zero for Iwi;::: wo/2 add high-
frequency energy to P*(w). For example, if F(wo) # 0, a tone with a
pitch equal to the sampling frequency will be heard. Gains as small as
1/100 or 1/1000 are not hard to achieve for Iwl ;::: wo/2. In many cases
the ear is not sensitive to the high frequencies and hence they are not
objectionable. At a sampling rate of 30,000 Hz, all high-frequency
distortions are at frequencies greater than 15,000 Hz and hence are
almost inaudible.
One other limitation of realizable filters must be taken into account.
They require a certain frequency band to change gain from unity to
zero. In Fig. 11, the transition occurs between We and wo/2. Large
distortions occur in this band; therefore it cannot contain useful
components in P*(w). We is effectively an upper limit for the usable
frequency of P*(w), which is less than the theoretical maximum wo/2.
Typically We = .8wo/2.
The spectrum P*(w) and hence p*(t) can be computed from Eq. 6 for
any smoothing filter F(w) and any pew). Thus the error pet) - p*(t) can
be computed. The calculation is complicated and is usually not worth
carrying out. Instead, either a physical feeling for the error is obtained
from a sketch such as Fig. 11 or bounds are computed for the error.

*Alternative Analysis of Sampling 2

In sampling, we measure the amplitude of some voltage function
pet), making the measurement R times per second. This sequence of R
measurements per second constitutes the samples of the waveform pet).
The process of successive measurements of the amplitude of pet) can
be carried out as shown in Fig. 12 by multiplying pet) by a succession of
R equally spaced impulses per second, each with unit area. Thus the
area (voltage times time) of each sample will be unity times the voltage

Voltage
funct ion
p (t)

Impulses
at rate R
per second

Fig. 12. Sampling-desampling process.

2 Suggested by J. R. Pierce. This analysis is briefer than the preceding one and
may be easier to understand.
 FUNDAMENTALS 17

of pet) at the time of sampling. The number in the computer represents

this area and is proportional to it.
We will disclose the consequences of such sampling by c~rrying out
the multiplication in the frequency domain. The function pet) will be
represented by its Fourier transform

pet) = LOO A(f) cos [27Tft - <p(f)] df

Here A(f) is the amplitude of the voltage spectrum at the frequency f,

and <p(f) is the corresponding phase.
The voltage Vs of the R unit-area impulses per second is

Vs = R( I + n~l 2 cos 27TnRt)

In the process of multiplication, each spectral component of Vs
interacts with the spectrum of pet) to produce a new spectrum. The dc
component of Vs produces simply pet) mUltiplied by R. Each other
spectral component of Vs produces a pair of sidebands lying about its
frequency, nwo (wo = 27TR). To see this, note that by elementary
trigonometry
(2R cos 27TnRt) cos [27Tf - <p(f)]
= R{cos [27T(nR - f)t + <p(f)] + cos [27T(nR + f)t - <p(f)]}
Thus if we plot the amplitude voltage spectrum of the sampled wave,
that is, of the samples, it appears as shown in Fig. l3. We see that if

Spectrum

R·A(fl R·A (f-R) R·A(f-2R)

,,------', /" .. ----- .... , ,---
, ,~ ,;(~
"A, '" ....
RI2 R 312R 2R

Frequency (Hz)

Fig. 13. Amplitude spectrum of sampled function.

pet) contains frequencies higher than R/2, that is, if A(f) is not zero for
f larger than R/2, the sideband lying below the sampling rate R will
fall partly within the frequency range from 0 to R/2. The higher
frequencies of pet) will have been folded over into the frequency range
from 0 to R/2.
18 CHAPTER ONE

Let us return to Fig. 12, which illustrates the sampling process. Here
we show the sampler (multiplier) followed by an amplifier of gain l/R
and a smoothing filter whose purpose is to remove frequencies above
R/2 Hz.
Suppose first that pet) contains no frequencies above R/2, and second
that the smoothing filter has zero loss for all frequencies below R/2
and infinite loss for all frequencies above R/2. Then from the preceding
analysis the output of the system should be exactly pet).
That ideal performance can fail in two ways.
The voltage function pet) may contain frequencies higher than R/2.
In that case, folded-over frequencies will appear in the frequency range
o to R/2, even though the smoothing filter is ideal.
The voltage function pet) may contain no frequencies higher than
R/2, but the smoothing filter may pass frequencies higher than R/2.
In that case, some folded-over frequencies above R/2 Hz will pass
through the smoothing filter.
In practice, we cannot make ideal smoothing filters. Rather, we count
on using frequencies only up to some cutofffrequency fc, which is some-
what less than R/2, and try to make the smoothing filter loss increase
rapidly enough with frequency above fc so that it passes little energy of
frequency above R/2.

Bounding Sampling Errors

In most cases of practical interest, only four parameters describing
the input spectrum P(w) and the filter transfer function F(w) need be
considered to characterize the sampling errors adequately. These
parameters are shown in Fig. 14.
The input spectrum (Fig. 14a) has been normalized to have unity
maximum magnitude. The maximum magnitude for w ;::: wo/~ is m.
The fold over error is characterized by the ratio m/l.O or simply m.
This number characterizes only the first foldover due to P(w - wo)
and neglects higher folds on the assumption that P(w) decreases rapidly
at higher frequencies.
Allowable values for m are not well known. Values of 2 or 3 percent
correspond to intermodulation distortions in much electronic equip-
ment. If the frequency at which the m peak occurs is near wo/2, large
values can be tolerated because the folded energy will be generated at
high frequencies (near wo/2). By contrast, if the peak occurs near wo, the
folded energy will appear at low frequencies and may be prominent and
objectionable.
 FUNDAMENTALS 19

\P(W)\

1.0

m
W
0 Wo

(a)

\F{w)1

1.0
-~I-C

0 We woll wQ

(b)

Fig. 14. Constants for bounding the error of the sampling process: (a)
spectrum of signal; (b) transfer function of :filter.

In trying to estimate whether the foldover of a given function pet)

will be objectionable, a mental comparison of pet) with certain known
waveforms may be useful. Figure 15 shows four waveforms-periodic
impulses, a square wave, a triangular wave, and a sine wave.
The impulse function (Fig. lSa) has harmonics that do not decrease
in amplitude; that is to say, the amplitude of the higher harmonics is
the same as that of the fundamental. Such a function has prominent and
usually objectionable foldover at any pitch. That behavior is typical of
any function containing sharp pUlses.
Figure lSb shows a square wave. Its harmonics decrease in amplitude
as 1jfrequency or 6 dB per octave. The square wave is usually good at
low pitches (pitch less than 200 Hz for a sampling rate of 10,000 Hz).
That behavior is typical for functions with sudden discontinuities.
Figure ISc shows a triangular wave. Its harmonics decrease as
Ij(frequency)2 or 12 dB per octave. It can usually be reproduced at
pitches up to 1000 Hz with a 10,000-Hz sampling rate. It is typical of
continuous functions with discontinuous derivatives.
20 CHAPTER ONE

....
Q. j-.JL-~'L....--.II----II--'-"L""'Time

(0)

( b)
Fig. 15. Examples of p(t)
functions with differing
foldover.

(c)

:;tf\
~V
f\ f\ !\
V V\TV
I. Time

(d)

To control the foldover of a very high-pitched function, pet) may be

formed from a sum of sine waves (Fig. 15d). In this way, foldover can be
completely eliminated simply by having no components above wo/2.
The number of components must be changed at different pitches, which
is an inconvenience.
Figure 14 shows the filter transfer function F(w). It has been normal-
ized to be approximately unity at low frequencies. The filter can be
specified to deviate from unity gain by no more than c at low frequencies
(w ::::; we) and to have a maximum gain (leakage) b at high frequencies
W ~ wo/2. The sharpness of frequency cutoff is measured by (wo/2) - We'

°
The usable frequency range is from to We; hence We should approach
wo/2.
Filter design and construction is a highly developed art. Typical
values that are easy to obtain in specially designed filters are c = .1
{l dB in-band deviation), b = 1/1000 (60 dB out-of-band attenuation)
and We = .8wo/2. General purpose filters or adjustable filters are not as
good but are more convenient to buy and use. It is always desirable to
have a flat in-band filter (c small). The importance of the out-of-band
attenuation depends on the sampling rate. At low rates (10,000 Hz),
 FUNDAMENTALS 21

out-of-band energy from 5000 Hz to 10,000 Hz must be carefully

removed. At high rates (30,000 Hz), the out-of-band energy is above
15,000 Hz and hence is almost inaudible. Hence some sloppiness in the
high-frequency filter performance is often tolerable. Note that such
leniency does not apply to the high-frequency parts of pew). Here, high
frequency energy folds and appears at low and prominent frequencies.

*Sample and Hold Analysis

The desampling process that we have analyzed assumed that impulses
or very narrow pulses were put into the smoothing filter. In actual
operation, wide pulses are usually used. A typical case is sketched in
Fig. 3b, where the pulse width is 80 percent of the sampling time T.
The gain of the desampling process is proportional to the pulse width,
hence the advantage of wide pulses. However, a small distortion which
we will now analyze is thus introduced in the spectrum of P*(w). The
distortion amounts to 4 dB in the worst case and is usually insignificant.
The holding process can be represented by introducing a filter
between the impulse modulator and amplifier in Fig. 6. The impulse
response of the filter h(t) is as shown in Fig. 16. Each impulse from the

t sec
w rad Isec

(0 ) (b)

Fig. 16. Sample and hold circuit: (a) impulse response; (b) frequency
function for D = T.
modulator is held for D seconds, thus producing a finite pulse. The
transfer function H(w) of this filter can be written

H(w) = CD ! e- jwt dt
Jo D
The amplitude of the impulse response is taken as liD to normalize the
low frequency gain of H(w) to unity. Carrying out the integration,
H(w) is evaluated as

H(w) = e -jwD/2. sinwD/2

wD/2
22 CHAPTER ONE

The maximum value of D is the full sampling time D = T. The

magnitude of H(w) for D = T is plotted in Fig. l6b. The maximum
distortion in P*(w) introduced by H(w) occurs at w = wo/2 and is 2/7T
or about 4 dB. The distortion decreases rapidly at lower frequencies.
If D = T /2, the maximum distortion which again occurs at w = wo/2
is only .9 or 1 dB.

*Analysis of Quantizing Errors

Quantizing errors are similar to noise in conventional electronic
equipment. The two most important characteristics of noise are its
magnitude and its frequency spectrum. We will derive an estimate of
these for quantizing errors.
These errors are shown graphically in Fig. 17. A pressure wave pet)
is sketched in Fig. 17a with a much enlarged ordinate, so that the

0
-I 0 2 3 4 5 6

(0 )
Time in
T units

o I I I I r
"l
-1/2
-I
I
0 2
1
3 4
I
5

(b)

Fig. 17. Quantizing process: (a) function being quantized; (b) quantizing
error.
quantizing levels 0, 1, and 2 are clear. The exact values p(iT) of pet)
at the sampling times are indicated by open circles. The analog-to-
digital converter approximates these by the nearest quantizing level
shown by the black dots pq(iT). The difference ei where
e1 = p(iT) - pq(iT) (8)
is the quantizing error.
 FUNDAMENTALS 23

A sketch of the error is shown in Fig. 17b. The maximu!Jl magnitude

of the error is -!- because of the nature of the analog-to-digital converter.
F or this analysis, we will assume that each ei is a random variable
uniformly distributed over the range --!- to +-!-. We will also assume
that ei's at different sampling times are uncorrelated, that is, that
°
E<eiej) = for i =I j. The notation E< ) denotes expectation taken by
an appropriate average.
These two assumptions are reasonable for pet) functions resulting
from almost all sound waves. They are the only assumptions that lead
to a simple estimate of the error. They neglect possible correlation
between pet) and the error. This correlation has been noticed in one
circumstance. DUring silent intervals, pet) = 0, and ei is constant. The
spectral energy of the quantizing error will be entirely at zero frequency
and thus inaudible. During nonsilent periods, the energy of ei will be
distributed across the audible spectrum. In this way the apparent noise
seems to fluctuate with the signal, disappearing during silent intervals.
This behavior contrasts with a normal tape recorder, in which back-
ground noise (tape hiss) is most prominent during silent intervals. For
many listeners, the quantizing noise is less objectionable than an
equivalent amount of tape hiss, because it tends to be masked by the
signal when it is present.
We will now continue with the error analysis. Without quantizing
errors the output of the sampIing-desampling process shown in Fig. 6
can be written
00

p*(t) = T L
i= - 00
p(iT)f(t - iT)

where f(t) is the impulse response of the smoothing filter and is related
to the filter frequency function by
1 f+oo
f(t) = -21T _ 00 F(w)e jwt dw

If the quantized samples pq(iT) are used as input to the impulse modu-
lator, then the output p:(t) is

p:(t) = T ~ pq(iT)f(t - iT)

i= - 00

The quantizing error eq(t) in the output is simply the difference

p*(t) - p:(t) and thus can be written
00

eit) = T L
i= - 00
{p(iT) - pq(iT)}f(t - iT) (9)
24 CHAPTER ONE

which by Eq. 8 becomes

00

eq(t) = T 2:
i= - 00
eif(t - iT)

The correlation function of eq(t) is by definition

. pq(T) = E(eq(t)eq(t + T»
Substituting Eq. 9 into the definition of pq( T), and taking advantage of
the noncorrelated et's (E( eiej) = 0 for i = j), the correlation function
simplifies to

, piT) = e 2 ·T f-+: f(t)f(t + T)dt

where ~ is the mean-square quantizing error

e2 = E(eiei)
The energy spectrum of the quantizing error is by definition the
transform of the correlation function
1 _
<I>q(w) = 27T f+ 00
00
.
pq( T)e- Jwr dT

and is

(10)

where F(w) indicates the conjugate of the frequency function of the

filter.
For ei uniformly distributed from -t to +t

- It
e2 =
-t
x2 dx = l-2
For the ideal smoothing filter, F(w) = 1 for Iwl < wa/2 and F(w) = 0
for Iwl ~ wa/2, the energy in <I>q(w) is uniformly distributed over the
frequency band - wa/2 to wa/2. The mean-square quantizing error

eit)2 = f:: <l>iw) dw (II)

W O/ 2 T-
=
f- wo/2
- e2 dw
27T

(12)
 FUNDAMENTALS 25

The spectrum of the quantizing error can be computed from Eq. 10

and the meari-square error from Eq. 11 or Eq. 12. Thus we have
completed our evaluation of quantizing error.
In order to compute the signal-to-quantizing-noise ratio, it is neces-
sary to specify the signal. For example, if the signal is a sinusoid that
occupies the entire range of quantizing levels (full-scale signal), it can
be written
N/2 sin wt

where N is the maximum number expressible with digits. The mean-

square signal is N 2 /8, and the ratio, rms signal to rms quantizing noise,
is

Earlier in the chapter we approximated this ratio as simply N/.5, which

is reasonable in view of the assumptions made in the analysis.
In general, the rough considerations of quantizing errors discussed in
the section on Sampling and Quantizing, are sufficient to control
quantizing errors. A more precise analysis can be done, as outlined
here, but is seldom worthwhile or necessary.

Digital-to-Analog and Analog-to-Digital Converters

Conversion between numbers in a computer and analog voltages is an
essential step in sound processing. Happily, it is conceptually simple
and practically easy to accomplish. A variety of commercial equipment
can be purchased. Complete converters come as a unit, or they can be
assembled from printed circuit cards sold by many computer companies.
The commercial units and the assembly techniques are described in
detail by their manufacturers; we will not reproduce this material, but
simply explain the way they work and point out some of the errors and
limitations.
Figure 18 shows the essential parts of a simple digital-to-analog
converter. A binary number can be expanded as the sum of its digits
times an appropriate power of 2. Thus, for example,

At the input to the converter, the five digits that make up the number
are represented by the voltages on five lines going to the switch controls
S4' .. So. A "I" is represented by a positive voltage and "0" by a
26 CHAPTER ONE

1/2n

I/lsn
Analog
output

'~----~~~~----~
Digital input

Fig. 18. Simple digital-to-analog converter.

negative voltage. The switch controls close their attached switch if they
have a positive input and open it with a negative input.
The resistor network embodies the sum given above. The resistors
are chosen to be inversely proportional to powers of 2. If Fi is a
switching function that is 0 if Si is open, and 1 if Si is closed, then
I = E R {F4·l6 + F 3 ·8 + F 2 ·4 + F 1 ·2 + Fo·l}
Thus I is the analog equivalent of the digital input. The constant of
proportionality is determined by the reference voltage ER • The current-
to-voltage amplifier generates an output voltage Eo which is proportional
to I.
In an actual converter, the switches would be transistors, the switch
controls would be flip-flop registers, the current-to-voltage amplifier
would be an operational amplifier, and the resistors would have values
measured in thousands of ohms. Higher accuracy and more digits are
obtained simply by adding more switches and resistors. Thus an
actual converter is not much more complicated than the simple device
we have described.
An analog-to-digital converter is more complicated. Most involve a
digital-to-analog converter plus a feedback mechanism, The exact
operation differs for different converters, but one widely used pro-
cedure is sketched in Fig. 19. The digital-to-analog converter that it
contains can be made in the way that has been described. The compli-
cated part is the programmer, which is effectively a small computer. A
conversion is made in a sequence of steps. The analog voltage to be
converted is applied to the analog input terminal. The programmer
initially sets all the digits S4' .. So equal to zero. Digit S4 is set to "1 "
 FUNDAMENTALS 27

Analog
,.. input

~
Digital- to-
analog
converter E2 "\
Comporer

h
~

Di9itol
~

,.. output
}
~

S4 S3 S2 S, So

Programmer

,
Fig. 19. Analog-to-digital converter.

as a trial. The resulting voltage E2 from the digital-to-analog converter

is compared with the input E 1 • If E2 ~ Eb then S4 remains "1." If
E2 > Eb S4 is reset to "0." The programmer carries out the same
process with S3' .. So. After five cycles involving five decisions from the
comparer, S4' .. So is the digital equivalent of the analog input.
What are the limitations on speed and accuracy of converters?
Usually speed is no problem with sound signals where sampling rates
need be no faster than 40,000 Hz. The basic limitation on speed is the
time for an electronic switch So - S4 to close, and for the transient
voltages produced by the switching to disappear. Switches that settle in
less than a microsecond are easy to build; hence sampling rates
approaching 1 MHz are routine for digital. to-analog converters.
Faster converters, up to 10 MHz, have been built using special circuits.
The analog-to-digital converter, as we have described it, is inherently
n times slower than a digital-to-analog converter, where n is the number
of digits. This limitation arises from the n sequential decisions involved
in converting a single number, each decision requiring a digital-to-analog
conversion. Thus, for example, a ten-digit converter with a I-flS
digital-to-analog part would have a maximum speed of 100 KHz.
One insidious error is inherent in the switching transients of a digital-
to-analog converter. If all the switches do not operate at exactly the
same speed, large errors will occur briefly during the change from certain
digits to adjacent digits. For example, in going from 0111 to 1000 the
analog output should change only one unit. However, all the digits
28 CHAPTER ONE

change state. If the most significant digit is slightly faster than the other
digits, the actual sequence will be 0111 1111 1000. The analog output
resulting from the correct and erroneous sequence is shown in Fig. 20.
It is clear that a large error is made momentarily. The error is difficult
to observe because it depends on the signal, that is, it depends on
transitions between particular levels, and it occurs very briefly.

Analog output

16 1111

8
~
0111
JL0111

O~-------------------------------------'Time

Fig. 20. Switching speed errors in digital-to-analog converter.

The error can be avoided in two ways. The switches can be carefully
adjusted to have the same operating speed. A good commercial con-
verter is usually satisfactory in this respect, whereas converters
assembled from computer cards may need adjustment. Secondly, a

Digitol-to- o ~--t----t--t----...
analog
converter
On
Off

Time

Fig. 21. Sampling switch to remove switching-speed errors.

 FUNDAMENTALS 29

sampling switch may be installed after the converter in order to gate

out the error pulses. Figure 21 shows the connection and a graph of the
operating waveforms.
Other errors in converters are fairly obvious. The quantizing error
due to the finite number of digits has already been discussed. Fourteen
is the maximum number of binary digits routinely available in com-
mercial converters; 12 digits are often used for sound.
The resistors in the network attached to the switches must be accurate,
particularly the small resistors. The tolerances can be maintained so that
the maximum resistor error is much less than the least significant digit.
Other errors are essentially standard signal-to-noise problems in-
herent in all amplifiers and electronic equipment. With proper design,
these can be kept insignificant in sound processing.

Smoothing-Filter Design
Filter design and construction is a highly developed science and art.
Satisfactory smoothing filters can be either built or purchased. They can
be of special design or of a standard type, or they can be variable with
knob-controlled cutoff frequency. Consulting a filter expert is the best
way to get just the right filter for a particular application. However, we
will give instructions for building one smoothing filter that has been
used for several years and is not too complicated.
The filter transfer function and circuit are shown on Fig. 22. 3 The
ver~ion shown is intended for a 20-KHz sampling rate. It has less than
1 dB loss over the band 0 to 8 KHz. It has 60 dB or greater loss for all
frequencies above 10KHz. The filter is not corrected for phase and will
distort the waveform of some signals. The phase change is less than that
introduced by any tape recorder and is almost always inaudible.
In constructing the filter, the components should be adjusted to be
within 1 percent of the values shown. An impedance bridge is used for
the adjustment. Capacitors can be adjusted by obtaining one that is
just under the desired value and adding a small capacitor in parallel.
Inductors can be adjusted by obtaining an inductor just larger than the
desired value and unwinding a few turns of wire. High-Q inductors of
good quality should be used, for example, those with torodial or ferrite
cores. The resistors are part of the source and load impedances and are
usually not built into the filter.

3 This filter was designed by F. C. Dunbar of the Bell Telephone Laboratories,

Murray Hill, New Jersey.
30 CHAPTER ONE

.107 .0917 .0658

5

t
5 eout

-10

- 20

m -30
"'0
c:
G)
..... -40
:::J
0
Q)
-50

-60

-70

-80
2
Frequency (KHz)

Fig. 22. Smoothing-filter circuit and transfer function. The filter has a
dc gain of 1-, which is not shown on the curve. Element values in KO, H,
and (1.f.

Filters for other sampling rates can be built from this design by
changing the values of the inductors and capacitors according to the
equations

C' = C· 20,000/fs
L' = L·20,OOO/fs

where C and L stand for the element values in the original design, C'
and L' stand for the element values in the frequency-scaled design, and
fs is the new sampling rate. For example, a 10-KHz sampling rate is
accommodated by doubling all inductors and capacitors.
As is shown on the circuit, the filter is designed to be driven by a
5-KG source impedance and to drive a 5-KG load. These impedances
are not critical. The source impedance may vary from 2 KG to 5 KG,
 FUNDAMENTALS 31

and the load impedance may vary from 5 KO to 100 KO without

seriously changing the transfer function.
A filter can be constructed in a day with about $50 worth of parts. It
is reasonably insensitive to the minor vicissitudes of a filter's life and has
generally provided satisfactory performance.

Digital Data Storage and Retrieval for Sound

The nature of samples of sound waves requires some special con-
sideration which will be developed here. A small amount of sound is
represented by very many numbers. For example, one minute of sound
sampled at 30 KHz produces 1,800,000 samples. When making either
an analog-to-digital or a digital-to-analog conversion, the samples must
be converted at an absolutely uniform rate! Variations in sampling
rate are equivalent to flutter or wow in an ordinary tape recorder and
are both audible and objectionable.
The number of samples is greater than the magnetic core memory of
most computers; hence the samples must be stored in some bulk
storage device. Fortunately since the samples are stored and retrieved
in sequence, a digital magnetic tape is ideal. 4 However, most digital
tapes do not store data continuously, but rather in groups called
records. In order to send the samples to the converter at a uniform rate,
a small core memory or buffer must be inserted between the tape and
the converter.
A typical digital tape is t-inch wide and 2400 ft long, and records
data on six tracks at a density of 800 digits per inch. Thus 400 12-bit
sound samples can be recorded on each inch. Allowing 10 percent of the
tape for record gaps, the entire tape will hold 107 samples or 300 sec of
sound sampled at 30,000 Hz. This is a practical if not large quantity.
The grouping of data into records is illustrated in Fig. 23. The record
gaps provide space to start and stop the tape. The record and playback

L~~~
~ LRecord gap
Fig. 23. Sample of
digital magnetic
tape showing
Record of data record gaps.

heads are initially positioned at the first record gap. The tape is started,
one record of data is transmitted, and the tape is stopped with the
4 Magnetic disk recording is also possible but has little advantage over tape
since the sound samples are in such an orderly sequence.
32 CHAPTER ONE

record-playback heads at the next record gap. The tape is kept in

motion through an entire record, since reading is unreliable during
starting and stopping.
The minimum length of the record gap is determined by the tape
speed and the rapidity with which it is started and stopped. A i-inch
gap and starting and stopping times of 3 to 5 msec are typical.
The maximum data rate from the tape is simply the tape speed times
the density of samples per inch. Thus typical speeds ranging from 60
inches per second to 150 inches per second and a density of 400 samples
per inch correspond to maximum data rates of 24,000 Hz to 60,000 Hz.
Achievable rates are slightly less than these maxima because of time
spent starting and stopping.
The control mechanism to start and stop a digital tape recorder, to
store the digital samples, and to transmit them to a converter at a
uniform rate is unfortunately complicated and expensive. A simple
schematic diagram is shown in Fig. 24. The digital tape transmits a

Digital-
to- Analog
analog output
converter
Put out a
Start-stop sample

Sampling-rate
oscillator

Fig. 24. Digital-tape control for sound.

record of data to the buffer memory. The buffer is a core memory

controlled so that the output samples will be in the same sequence as
were the samples put into it. It has sufficient flexibility so that its output
and input can be interleaved in any order. The sampling rate is, and
should be, controlled by a single oscillator, which can easily be set to any
desired sampling rate. Each cycle from the oscillator causes the buffer
to deliver another sample to the converter and the converter to output
the sample. The control circuits keep track of the number of samples in
the buffer memory and start the tape recorder before it is empty.
The size of the buffer memory is determined by the record length of
the tape data. If any sampling rate from zero up to nearly the tape data
rate is to be accommodated, then a buffer longer than one record is
necessary. A length equal to two records simplifies the control circuits.
 FUNDAMENTALS 33

The design must take consideration of these facts: larger buffers cost
more, longer records yield higher maximum rates because of fewer
record gaps, the tape must be started soon enough to avoid emptying the
buffer at the highest sampling rate, the buffer must be large enough
never to overfill at the lowest sampling rate. A design is a proper com-
promise between these factors. Although we have only discussed a
digital-to-analog conversion system, the analog-to-digital process
requires the same buffer and works in a basically similar manner.
The digital tape controller that we have described is rather expensive
and complicated to build. Often the computer itself makes a more
attractive tape controller. A schematic diagram is shown in Fig. 25.

External
data
connection

Computer Digital-
plus to- analog
program ....+-----4-..... converter

Sampling- rate
oscillator

Fig. 25. Computer as tape control for sound.

In order to function in this way, the computer must have an external

data connection which will deliver samples to a converter under the
control of an external oscillator. Most recent computers can be obtained
with this feature. The digital tape transport can be one normally
associated with the computer, the computer's core memory serves as
buffer, and functions of the control circuits are accomplished by a
suitable program. Thus the same machine that synthesizes or analyzes
the sound can also communicate directly in sound with the external
world.

Fundamental Programming Problems for Sound Synthesis

In the preceding material we have described a powerful and flexible
technology for sound processing by computer. The remaining ingredient
is the computer programs that activate this technology; but that is a
large ingredient. Most of the rest of this book can be considered as
descriptions of some of these programs. Sound processing can be
divided into sound analysis and sound synthesis. So far, no universal
34 CHAPTER ONE

programs for analysis have been developed. Rather, many different pro-
grams have been written for particular tasks. For synthesis, one
program, which developed through five stages, Music I-Music V, has
proved generally useful. Hence we will present here the fundamental
considerations that led to Music V, and in the next chapters details
intended to teach a user of Music V. However, the material should be of
value not only to users of Music V, but to anyone writing a sound-
synthesis program.
The two fundamental problems in sound synthesis are (1) the vast
amount of data needed to specify a pressure function-hence the
necessity of a very fast and efficient computer program-and (2) the
need for a simple, powerful language in which to describe a complex
sequence of sounds. Our solution to these problems involves three
principles: (1) stored functions to speed computation, (2) unit-generator
building blocks for sound-synthesizing instruments to provide great
flexibility, and (3) the note concept for describing sound sequences. Let
us next consider sound synthesis from the computer's and the
composer's standpoints to see the importance of these principles.
To specify a pressure function at a sampling rate of 30 KHz, one
number is needed every 33 fLsec. That speed strains even the fastest
computers. A useful measure of computation is the time scale, which is
defined as
TO I _ time to compute samples of a sound
lme sca e = duration of the sound
Various possibilities exist at various time scales. If the time scale is
equal to 1 or less, a digital-to-analog converter can be attached directly
to the computer and sound can be synthesized in real time. This allows
improvising on the computer, hearing the sound as one pushes the
computer keys in the same way that one hears sound from a piano.
Fast current computers add two numbers in about 3 fLsec and multiply
two numbers in about 30 fLsec. Hence the computations for each
sample for real-time synthesis must be few indeed. However, real-time
synthesis is a powerful way of adjusting sound parameters to achieve
a particular timbre or effect. In addition, it allows the computer to be
used as a performing instrument. Hence it is an important objective.
Time scales greater than 1 necessitate recording the samples on a
digital magnetic tape, rewinding the tape, and playing the tape through
the converter. A delay equal to or greater than the sound duration is
inherent in the process. Time scales from 1 to 50 are eminently usable.
At 50, a delay of an hour is needed to compute one minute of sound.
 FUNDAMENTALS 35
An hour seems long if you personally are waiting for the computer; it is
nothing if you are at home sleeping while the night shift runs the
problem. At a scale of 50, 1600 fJ-sec are available to compute each
sample. Fifty multiplications or several hundred additions can be
carried out in that time. Although much can be done, that number of
computations does not represent a copious supply, and it must be used
effectively.
Time scales from 50 to 1000 become so time consuming and expensive
that even the most reckless experimenter pauses to consider whether
the value of his sounds justifies the time and money. At a scale of
1000,20 minutes of computer time are needed for each second of sound.
It must be a remarkable second to make this effort seem worth while.
One way of speeding the effective computation is to store samples in
the computer memory, when possible, and to read these samples from
memory rather than recompute them. Reading from memory is rapid.
The process works only for samples or factors of samples that are
repetitive. Fortunately, many sounds have highly repetitive com-
ponents. For example, an oscillator repeats the same waveform each
cycle. The shape of a cycle~s wave may be very complicated, but once it
is computed and stored, it can be read out as rapidly as any simple
function. Many other factors can be reduced to repetitive stored
functions.
The cost of stored functions is memory space. In Music V a typical
function is stored as 512 samples, and the largest part of the memory
is used for storing functions. The cost is more than justified by the time
saved.
We have considered sound synthesis from the position of the
computer and it has led us to stored functions. Now let us look from
the composer's standpoint. He would like to have a very powerful and
flexible language in which he can specify any sequence of sounds. At
the same time he would like a very simple language in which much can
be said in a few words, that is, one in which much sound can be
described with little work. The most powerful and universal possibility
would be to write each of the millions of samples of the pressure wave
directly. This is unthinkable. At the other extreme, the computer
could operate like a piano, producing one and only one sound each time
one of 88 numbers was inserted. This would be an expensive way to
build a piano. The unit-generator building blocks make it possible for
the composer to have the best of both of these extremes.
With unit generators the composer can construct, with a simple
procedure, his own sound-synthesizing program. In Music V it is called
36 CHAPTER ONE

the orchestra, and it contains a number of different subprograms

called instruments. The unit generators perform functions that ex-
perience has shown to be useful. For example, there are oscillators,
adders, noise generators, and attack generators. Many unit generators
perform conceptually similar functions to standard electronic equipment
used for electronic sound synthesis.
In a given instrument the composer can connect as many or as few
unit generators together as he desires. Thus he can literally take any
position he chooses between the impossible freedom of writing indi-
vidual pressure-function samples and the straightjacket of the computer
piano. In this way, in unit-generator building blocks, we have given the
composer almost ultimate flexibility to choose the environment in
which to work. The price is the work he must do in constructing the
instruments in his orchestra. However, the language with which the unit
generators are assembled is so elegant that this cost is insignificant.
The final principle for specifying sound sequences is the note concept.
Sound exists as a continuous function of time starting at the beginning
of a piece and extending to the end. We have chosen, for practical
reasons, to chop this continuous sound into discrete pieces, called
notes, each of which has a starting time and a duration time. This
division is admittedly a restriction on the generality of sound synthesis,
but one we are not brave enough to avoid. Needless to add, notes have
been around for some time.
The note concept interacts with the instrument in a straightforward
way. The instruments are designed to "play" notes. At the starting
time of each note, a set of instructions is given to the instrument, and it
is turned on for the duration of the note. No further information is given
to the instrument during the course of the note; the complexity of the
instrument determines the complexity of the sound of the note.
The instructions for the instrument for each note are written on a
score by the composer (or by the composer's program if he wishes to
delegate this task to the computer). Hence, the complexity and length
of the instructions, multiplied by the number of notes, determines the
amount of work the composer must do. In general, complicated instru-
ments require more instructions, but they may be able to play longer
and more interesting notes. Within the limitations of these conflicting
factors, the composer must create an environment in which he is
willing to work.
The note concept also includes the idea of voices, which have their
usual musical meaning. In Music V instruments can play any number of
notes at the same time. The program adds all voices and puts out the
 FUNDAMENTALS 37

combined sound. The addition is simple for the computer. Furthermore,

it automatically synchronizes all voices. Each note has a starting time.
The computer arranges all notes in a composition into the proper time
sequence, and thus the composer can write the score in any order he
chooses. In this way the tyranny of time, which so harasses the per-
forming musician, is almost completely eliminated by the computer.
The fundamentals of stored functions, unit generators, and notes
have been given general consideration here. Details of their use in the
Music V program appear frequently in the following chapters.

Overview of Sound-Synthesis Program-Music V

Next we discuss the over-all operation of the Music V program, both
as an example of a sound-synthesis program and as an introduction to
the more detailed material that follows.
An outline of the program is shown in Fig. 26. Programs change; the
description given here is the program as it was created in 1967-1968.

Pass I: Pass II Pass m Sound

Score composer time sound
sequence sequence generation samples

Fig. 26. Outline of sound-synthesis program.

It ran on a General Electric 635 computer but was programmed almost

entirely in F0RTRAN IV to simplify its transposition to other machines.
F or concreteness, we will speak of this specific historical program.
However, most of what we shall say applies to other computers. Also,
the basic program has been modified in many ways for special purposes,
such as adding another input section to accommodate graphical scores.
These will be described later.
The composer speaks to the computer through a score that contains
not only the notes to be played but also descriptions of the instruments
on which they will be played.
The orchestra description specifies each instrument in the orchestra
in terms of the type of each unit generator in it and how the unit
generators are interconnected or related. Types of unit generators-
oscillators, adders, random sources-are straightforward. Many
interconnections are possible. For example, the outputs of two oscil-
lators can be added to produce a more complex tone, or one oscillator
can control the frequency of a second oscillator to produce a vibrato.
38 CHAPTER ONE

Inherent in the description of each instrument are the input parameters

needed to run it. For example, if the instrument is to play notes of
differing pitches, one input parameter must specify pitch. If vibrato is
to have a controllable rate, a parameter must specify this rate. These
parameters must eventually be supplied by the score.
Different instruments must be uniquely designated. This is done
simply by numbering them. Thus the program will have instruments
1, 2, 3, and so forth, and the score will request a note to be played by an
instrument number.
The sound synthesis is divided into three sections for both conceptual
and computational reasons. Pass I reads the score written by the
composer. It may contain note ·cards requesting the synthesis of a
specific note, instruction cards which cause Pass I to produce note cards,
and cards to set functions and parameters in Passes I, II, and III. Each
note card must contain an instrument number, the starting time of the
note, and the duration of the note. All other quantities on the card
depend entirely on what input parameters the composer has specified
in his instrument.
If notes are to be played, the score must also contain cards defining
instruments. These also contain the time at which the instrument is to
be defined. It is possible to redefine an instrument part way through the
composition, thus changing the timbre.
Pass I reads and processes the score in the sequence in which the
composer has written it. Note cards written by the composer are simply
sent directly to Pass II. Instruction cards may cause subroutines in
Pass I to generate or compose note cards which are then sent to Pass II.
Thus, most of the composing power of the program resides in Pass I.
The note cards written by either the composer or Pass I can be written
in any time sequence; this gives great flexibility. Furthermore, notes for
different voices can be intermixed in any order.
In Pass II all note cards are sorted into increasing time sequence to
prepare for sound generation. Parameters on the note cards may be
modified by F0RTRAN subroutines. For example, the frequency ratio
between two voices at a given time can conveniently be adjusted because
all voices are in proper time sequence. Metronome markings to acceler-
ate or retard the time scale are convenient to apply in Pass II. However,
once the notes are ordered for time, new notes cannot be added without
destroying the ordering. Hence, new note cards cannot be generated in
Pass II.
Pass III reads the note cards after they have been time otdered by
Pass II. At the beginning of each note, the parameters from the note
 FUNDAMENTALS 39

card are inserted into the appropriate instrument and the instrument is
turned on for the duration of the note.
To summarize, the complete program with three passes, stored
functions, unit generators, and instruments was evolved over several
years. It is not a unique way of synthesizing sound samples; other
equivalent programs could be written. However, it does provide great
speed and great flexibility by the careful use of a general compiling
language (F0RTRAN) plus certain machine language subroutines.

Annotated References by Subject

Sound in General
P. M. Morse, Vibration and Sound (McGraw-Hill, New York, 1948). A mathe-
matical, technical book written from the physicist's standpoint.
C. A. Taylor, The Physics of Musical Sounds (English Underwriters Press, London,
1965). A less technical book than Morse's; still from the physicist's standpoint.
Sampling, Foldover Errors, etc.
J. R. Ragazzini and G. F. Franklin, Sampled-Data Control Systems (McGraw-Hill,
New York, 1958). The first chapters treat sampling and desampling in general
and are not limited to control systems.
Quantizing Errors
W. R. Bennett, "Spectra of Quantized Signals," Bell Sys. Tech. J. 27, 446 (1948)
The fundamental analysis of quantizing errors. '
Ana log-Digital Conversion
B. W. Stephenson, Analog-Digital Conversion Handbook (Digital Equipment
Corporation, Maynard, Mass., 1964). A combined catalog and instruction
book for making digital circuits from plug-in components (manufactured
by the Digital Equipment Corp.).
Note: Several companies manufacture analog-digital conversion equipment.
These may change from year to year. Currently, the Texas Instrument
Company and the Raytheon Company make satisfactory apparatus for
sound processing.
Filter Design
M. E. Van Valkenburg, Introduction to Modern Network Synthesis (John Wiley &
Sons, New York, 1960). A good presentation of some of the many details of
network synthesis from the standpoint of the electrical engineer.
Digital Data Storage and Retrieval
E. E. David, Jr., M. V. Mathews, and H. S. McDonald, "Description and Results
of Experiments with Speech Using Digital Computer Simulation," Proceed-
ings of 1958 National Electronics Conference, pp. 766-775.
E. E. David, Jr., M. V. Mathews, and H. S. McDonald, "A High-Speed Data
Translator for Computer Simulation of Speech and Television Devices,"
Proceedings of I.R.E. Western Joint Computer Conference, pp. 354-357
(1959).
These articles describe some of the first equipment for computer sound processing.
The equipment is obsolete, but the principles are valid and important.
Fundamental Programming Problem
M. V. Mathews, "An Acoustic Compiler for Music and Psychological Stimuli,"
Bell Sys. Tech. J. 40, 677-694 (May 1961).
40 CHAPTER ONE

M. V. Mathews, "The Digital Computer as a Musical Instrument," Science 142,

553-557 (November 1963).
J. R. Pierce, M. V. Mathews, and J. C. Risset, "Further Experiments on the Use of
the Computer in Connection with Music," Gravesaner Blatter, No. 27/28,
92-97 (November 1965).
These are the original papers that trace the development of the current sound-
synthesizing program.

Problems for Chapter 1

Numerical Representations of Functions of Time
1. Sketch pressure functions that
(a) Are periodic with a period of 1 millisecond (msec)
(b) Have a fundamental pitch of 600 Hz
(c) Have energy only at 500 Hz
(d) Have energy at 500 Hz and 750 Hz (what is the period of this
function ?)
(e) Have energy only at zero frequency
(f) Have no perceivable periodicity
2. On the same sheet of paper, draw three functions that are periodic
with periods of (a) 10 msec, (b) 5 msec, and (c) 1 msec. What are the pitches
of these functions?
3. Desample the following sequences of numbers. Draw a graph of the
pulses from a sample-and-hold desampling circuit. Assume a 50-percent
duty factor for the pulses. Pass a smooth waveform through the pulses in a
manner in which you imagine a smoothing filter would operate
(a) 0,3,5,7,10,13,15,17,20,22,25,28,30,24,20,15,10,6,0, -4, -7,
-9, -10, -10, -9, -7,0
(b) 0, 7, 10, 7, 0, -7, -10, -7, 0
(c) 10, -10, 10, -10, 10, -10, 10, -10 (for case c draw at least two
possible smooth waveforms; which one would be passed by a low-pass
filter having a cutoff frequency appropriate to the sampling rate?)

Time (msec)

4. Sample and quantize 2 msec of the waveform shown above at a

sampling rate of
(a) 1000 Hz (b) 2000 Hz (c) 5000 Hz
 FUNDAMENTALS 41

(d) Which sampling rate seems "visually" sufficient to characterize the

waveform?
(e) Assume the wave has energy at only two frequencies; what are these
frequencies?
(f) What is the minimum faithful sampling rate according to the sampling
theorem?
(g) What is the period and pitch of the wave?

Sampling
5. A waveform p(t) where
p(t) = 100 sine (27T·2100·t) + 50 sine (27T·4200·t)
+ 33 sine (27T' 6300· t) + 25 sine (27T' 8400· t)
+ 20 sine (27T·I0,500·t) + 17 sine (27T·12,600·t)
+ 14 sine (27T·14,700·t) + 12 sine (27T·16,800·t)
+ 11 sine (27T·18,900·t) + 10 sine (27T·21,000·t)
is subjected to a sampling and desampling process as shown in Fig. 6.
The sampling rate is 19 kHz. The desampled output is p*(t).
(a) What is the highest frequency component that can be faithfully
reproduced in p*(t) at this sampling rate? Call this component and all
lower components the desired components. Give the amplitudes and
frequencies of components of p*(t) with
(b) no smoothing filter
(c) a filter with the frequency function

+3
I
I
I
I
0 I
I
I
I
I
ID I
"CI I
I
I
-20 ----------------~----
I
I
I
I
-40
0 8 9 20
KHz

(d) Also give the amplitudes and frequencies of the components of p(t).
(e) What is the lowest frequency component in p(t)? What is p(t)'s
period? What is its pitch?
(f) What is the lowest frequency component in p*(t) with no smoothing
filter? What is p*(t)'s period?
(g) How much are the desired components in p*(t) changed by the
filter? Which desired component is most changed?
(h) Which distortion components in the "range of perception" (0-15
kHz) are reduced by the filter? Which are relatively unaffected?
42 CHAPTER ONE

(i) What distortion components that are folded about 38 kHz fall in the
range of perception?
(j) With no filter, what is the maximum frequency that can be reproduced
without causing a distortion component in the range of perception?
(k) With the filter, what is the maximum frequency that can be repro-
duced without causing a distortion component in the range of perception
(assume filter has infinite attenuation for frequencies greater than 9 kHz)?
6. Samples at a 20-kHz rate are computed for the waveform

. [21T • (60,000.
f( t) = sme 30 t) . t ]

. [2 (30,000(20
+ sme 1T. 30 - t») . t]
The sound is desampled with an impulse desampler and no filter. Describe
the amplitudes and frequencies of the components that fall within the
range of perception (0 to 15 kHz); t goes from 0 to 30 sec.

Analog-Digital Conversion
7. Calculate the tolerance on the resistors in the digital-to-analog
converter shown in Fig. 18 so that the maximum error due to anyone
resistor is to of one quantizing level. Give tolerance in terms of both
absolute accuracy and percent accuracy. Which resistor must have the best
percent accuracy?

Smoothing Filter Design

8. Frequency scale the filter shown in Fig. 22 so that the cutoff frequency
is
(a) 15 kHz corresponding to a sampling frequency of 30 kHz;
(b) 3 kHz corresponding to a sampling frequency of 6 kHz.

Digital Data Storage and Retrieval

9. Suppose that you have a digital tape with data recorded at 400 samples
per inch and a tape speed of 100 inches per second. The record gaps are
1 inch long; the tape takes 6 msec to stop after reading the last sample in a
record and 10 msec to start (time from start signal to reading of first sample
in record).
(a) Calculate the maximum data rate for record lengths of 100, 500,
2500, and 10,000 samples.
(b) Calculate the minimum buffer size for each record length to accom-
modate data rates for 0 up to the maximum.
For safety, design the control so the buffer will never have less than 50
samples and will always have 50 or more empty cells.
(c) At what number of samples in the buffer should the tape start signal
be given?
2 A Sequence of Tutorial Examples
of Sound Generation

Introduction
This chapter is intended to provide a training course in the use of
Music V by discussing a series of examples ranging from simple to
complex sound synthesis. It is written from the point of view of the
user of Music V. Details of operation of the programs will be suppressed
as much as possible. These can be found in Chapter 3. Because the
programs will not be described here, many of the conventions of the
computer score will seem arbitrary and must be temporarily accepted
on faith.
For concreteness we will also arbitrarily assume values for certain
parameters of the program, for example, a sampling rate of R =
20,000 Hz. Other parameters will be introduced as required. For the
student's benefit, the parameters of the training orchestra are listed at
the beginning of the problems for Chapter 2.
The material assumes that the student has a working knowledge of
F0RTRAN programming. The programming examples will be written
in F0RTRAN IV. It is also assumed that the student understands the
general functioning of a computer-arithmetic, memory, input-output,
and program. If necessary, these skills can be learned from books cited
in the references at the end of Chapter 2.
This chapter is intended as training material and not as a reference

43
44 CHAPTER TWO

manual for Music V. Reference material is organized and presented in

Chapter 3.

The Simplest Orchestra

By way of introduction, an elementary orchestra and score are shown
in Fig. 27. We shall start by describing the material and then explain
the details of its operation. Figure 27a shows the conventional score of
the few notes that will be synthesized. Figure 27b shows the block
diagram of the simulated instrument that will play the score. It consists
solely of an oscillator and an output box. The oscillator has two inputs;
the amplitude of the output equals P5; the frequency is proportional to
P6. The waveform of the oscillation is determined by stored function
F2, which is sketched in Fig. 27c.
The records in the computer score, Fig. 27d have been numbered for
reference in this discussion. Each record has a sequence of entries
designated PI, P2, P3, etc. In the training orchestra, up to 30 entries
(PI-P30) may be used. The entries are separated either by blank spaces
or by a comma. Each record is terminated by a semicolon. A record may
extend over several lines ; conversely, several records may be put on one
line.
Records 1 through 4 define, for the computer, the instrument shown
in Fig. 27b. INS 0 I ; says that, at time 0 in the composition, instrument
1 will be defined. 0SC P5 P6 B2 F2 P30 ; says that the first unit generator
in the instrument will be an oscillator, will have inputs P5 and P6,
will use function F2 for its waveform, will store its output in 1-0 block
B2, and will use P30 for temporary storage (which we will discuss'later).
0UT B2 Bl; says to take the samples in 1-0 block B2 and add them to
the contents of block BI in preparation for outputting these samples.
END; terminates the instrument definition.
Record 5 defines the function F2 (Fig. 27c) and causes it to be
generated and stored in the computer memory assigned to F2.
Notes 1 through 11 in the score are generated by records 6 through
16, respectively. In each of the records PI (N0T) says the purpose of
the record is to play a note. P2 gives the starting time of the note
measured in seconds from the beginning of the composition. P3 (1)
gives the instrument number on which the note will be played. P4 gives
the duration of the note in seconds. Durations of staccato notes are
written to produce more silence between successive notes than the
corresponding silence for the legato notes. P5 gives the amplitude of
the note as required by the instrument. In the training orchestra, ampli-
tude can vary over the range 0 to 2047. Amplitudes are varied to
 A SEQUENCE OF TUTORIAL EXAMPLES 45

P5 P6

~
c
IJoe! pj l:~ J J 1;.1 J tr F2

~ ~-~-======= ===-
82

~UT
(a)
(b)

F2

511

(c)

1 INS 0 1 ;
2 0SC P5 P6 B2 F2 P30 ;
3 0UT B2 Bl ;
4 END;
5 GEN 0 1 2 0 0 .999 50 .999 205 - .999 306 - .999 461 0 511 ;
6 N0T 0 1 .50 125 8.45 ;
7 N0T .75 1 .17 250 8.45 ;
8 N0T 1.00 1 .50 500 8.45 ;
9 N0T 1.75 1 .17 1000 8.93 ;
10 N0T 2.00 1 .95 2000 10.04 ;
11 N0T 3.00 1 .95 1000 8.45 ;
12 N0T 4.00 1 .50 500 8.93 ;
13 N0T 4.75 1 .17 500 8.93 ;
14 N0T 5.00 1 .50 700 8.93 ;
15 N0T 5.75 1 .17 1000 13.39 ;
16 N0T 6.00 1 1.95 2000 12.65 ;
17 TER 8.00 ;

(d)

Fig. 27. Elementary orchestra and score: (a) conventional score; (b)
instrument block diagram; (c) waveform; (d) computer score.
46 CHAPTER TWO

correspond to the dynamic markings on the conventional score. P6

equals .02555 times the frequency of the note in cycles per second
(hertz). The proportionality constant .02555 will be explained below.
Record 17 terminates the composition at 8 sec.
Simple Unit Generators to Output, Add, and Multiply
Having introduced a simple orchestra and score from the user's
standpoint, we will now describe in more detail the operation of a few
simple unit generators. Although they are simple, these are the most
frequently used building blocks for all instruments.
As we showed in Chapter 1, the acoustic output wave is produced by
passing a sequence of numbers (samples) so, Sb .•• , S1 through a digital-
to-analog converter and driving a loudspeaker with the analog voltage
from the converter. The first sample So is the amplitude of the acoustic
wave at the beginning of the composition at t = 0, where t is time. The
second sample S1 is the amplitUde one sampling time later. We shall
assume a sampling rate of 20,000 Hz for the training orchestra; hence
S1 is put out at t = 1/20,000 sec. S40,000 is the amplitude at t = 2 sec.
It is quite possible, though seldom useful, to specify the sample that
controls the amplitude of the acoustic output at any 1/20,000 sec
throughout the entire composition.
The purpose of the portions of the Music V program called "instru-
ments" is to calculate all the S1 samples. For example, if a note is to be
played from 3 sec to 4 sec in the composition, samples S60,000 through
sao,ooo must be computed. The nature of S60,000 through sao,ooo deter-
mines the characteristics of the sound-its pitch, loudness, timbre,
everything. The nature of the samples is, in turn, determined by the
particular unit generators that are put together to form the instrument
and by the numbers on the data records that control these generators.
A problem that must be solved by Music V is to keep track of time
so as to "turn on" a given instrument program at the sample at which
its note should begin, and to "turn off" the instrument at the sample at
which its note shOUld end. The starting sample is computed simply by
multiplying the starting time of the note given in P2 by the sampling
rate. The terminating time is P2 plus the duration P4, and the ter-
minating sample number is the sampling rate times the terminating
time. Because of the universal necessity for this control of time, P2 and
P4 must always be used for starting time and duration in all records
which specify notes.
A second problem facing Music V is to combine the numbers from all
instruments that are playing simultaneously at a given time. The
 A SEQUENCE OF TUTORIAL EXAMPLES 47

digital-to-analog conversion process demands that the samples be

output in sequence, Sl followed by S2, followed by S3, and so forth. Thus
the contribution of all instruments to a sample must be computed
simultaneously. A way to accomplish this end, which has been used in
earlier programs, is to calculate one number from each active instru-
ment, combine these numbers (by addition), output the sample, and
then proceed to the next sample. Music V operates in essentially this
way, but for additional efficiency it calculates a block of numbers from.
each instrument instead of a single number. These blocks, called 1-0
blocks, are one of the fundamental data storage units in the program.

I-@ Blocks
1-0 blocks are short for unit generator input-output blocks. They
can be used as storage locations for either inputs or outputs for unit
generators, hence the designation input-output blocks. Blocks are
designated B1 through B lOin the training orchestra. Block B I has the
special function of storing the numbers that will be sent to the digital-
to-analog converter. All other blocks are equivalent in mono. (In
stereo, blocks BI and B2 are both reserved for output.)
The size of the block is a parameter of the orchestra. In the training
orchestra, it has been set at 512. The maximum size of numbers in the
I -0 blocks is another program parameter. In the training orchestra it
has been set at ± 2047 which is appropriate for a I2-bit digital-to-analog
converter.

AD2 Generator
The simplest generator is the two-input adder, AD2. Its function is to
combine two numbers by addition. It has two inputs and one output
as shown in Fig. 28a. The equation of operation is

where 11 and 12 are the two inputs, 0 is the output, and i is the index of
samples that starts at 0 at time t = O. We must quickly add that this
equation is computed only for those samples during which the
instrument with AD2 is playing a note.
In the score, AD2 is put in an instrument by a statement such as
AD2 B2 B4 B3 ;
This example says: take the numbers stored in block B2, add them to
those stored in block B4, and put the sum in block B3. The relation
48 CHAPTER TWO

II 12 II

¥ ~
(0 )
G (b)

II 12 II I2

~ ~

(e)
~ Fn

~
(d)

Fig. 28. Four simple unit generators: (a) AD2; (b) 0UT; (c) MLT;
(d) 0SC.

between sample index i and the numbers in a given block at a given time
need not worry the user; it is treated automatically by the program.
AD3 and AD4 also exist and form a sum of three and four inputs,
respectively. The score statement evoking AD4 would be
AD4 B2 B3 B4 B5 B6 ;
where B2 through B5 are inputs and B6 the output.

@UT Generator
The 0UT generator takes the numbers from an instrument and
places them in the special 1-0 block BI for subsequent outputting
through the digital-to:-analog converter. 0UT also combines the
numbers with any other instrument simultaneously being played. 0UT
is diagrammed in Fig. 28b. It is shown with one input. The output to
B I is not shown; it always goes to this block. The equation of operation
is in F0RTRAN-like nptation
Acoustic output! = acoustic output! + II!
This equation says: I I is added to anything previously in the acoustic
output block; by this simple means any number of instruments may be
combined. The operation of addition is perfectly equivalent to the
 A SEQUENCE OF TUTORIAL EXAMPLES 49

way in which sound waves of several real instruments combine in the

air.
In the score 0UT is evoked by a statement such as
0UT B2 B1;
where B2 is the block containing the input, and B1 is the special block
for acoustic output.

MLT Generator
The MLT generator multiplies two numbers together in a manner
exactly analogous to the addition done by AD2. It is diagrammed in
Fig. 28c. The equation of operation is
0 1 = 11 1 .121
where 11 and 12 are the two inputs and 0 is the output. In the score
MLT B2 B3 B4;
associates II with B2, 12 with B3, and 0 with B4. In general, the order
of listing generator descriptions on the score is: inputs, outputs, special·
parameters.

@sc Generator 1
By far the most important generator is the oscillator 0SC. It is the
most frequently used and the most difficult to understand of the simple
generators. Its importance is based on the prominence of oscillations in
musical sounds and on its nature as a source of numbers. The generators
previously described modify or output numbers that have been created
elsewhere; 0SC is one of the few units that actually produce numbers.
The diagram of 0SC is presented in Fig. 28d. As will be shown,
three quantities determine the output 0: 11 controls the amplitude of the
oscillation; 12 controls the frequency; and Fm a stored function, is the
waveform. Fn is exactly one cycle of the 0SC output; the purpose of
the 0SC can be looked upon as repeating Fn at the desired frequency
and amplitude.
Fn may be thought of as a continuous function of time, but in the
computer it must be represented by a block of samples. In the training
orchestra each function is represented by 512 samples. Figure 29 shows
an example of a stored function F3. The waveform is, a square wave
with slightly slanted sides. The 512 points, F3(k) k = O... 511, are

1 Also see Chapter 3, Section 6, for a basic discussion of 0SC.

50 CHAPTER TWO

F3(0) =0 F3(250) = 1.0 F3(505) = -1.0

F3(l) = .2 F3(251) = 1.0 F3(506) = - 1.0
F3(2) = .4 F3(252) = .8 F3(507) = -.8
F3(3) = .6 F3(253) = .6 F3(508) = -.6
F3(4) = .8 F3(254) = .4 F3(509) = -.4
F3(5) = 1.0 F3(255) = .2 F3(510) = -.2
F3(6) = 1.0 F3(256) = 0 F3(51l) =0
F3(257) = -.2
F3(258) = -.4
F3(259) = -.6
F3(260) = -.8
F3(261) = - 1.0
F3(262) = - 1.0

Fig. 29. Function stored as 512 samples.

indicated as dots on the function. Actually only 511 numbers are

independent since F3(O) = F3(511). The 512 numbers representing the
function are listed below the function. These numbers are actually
stored in 512 locations in the computer memory. The programs that
calculate and store the numbers are called GEN routines and will be
discussed later.
One may ask, why go to all the trouble of having a GEN program
compute and store numbers and then have the 0SC program modify
and repeat these numbers? Why not, instead, have the GEN programs
repeatedly calculate exactly the desired numbers? The reason, the
importance of which cannot be overemphasized, is efficiency. 0SC is a
very fast number repeater. The GEN programs-must be flexible and,
hence, they are in comparison very slow.
By denoting a function F3, we imply that several stored functions are
possible. In the training orchestra 10 functions, designated Fl through
FlO, are available.
 A SEQUENCE OF TUTORIAL EXAMPLES 51

The simplest 0SC program would simply repeat the 511 numbers in
F3, one after the other: F3(0), F3(l), ... , F3(511), F3(1), .... This
would produce an oscillation whose peak amplitude would be 1 and
whose frequency would be 20,000/511 = 39.14 Hz. That frequency is
too low for most purposes. By repeating every other sample, F3(1),
F3(3), ... , F3(511), F3(2), ... , one could produce a higher frequency,
78.28 Hz. In general, by repeating every nth sample of F3, one obtains
a frequency of

20,000 H
----sfl.n z

F3 is stored as samples, as is the output of 0SC: the process carried

out by 0SC can be thought of as resampling F3 to obtain a desired
frequency. A simple resampling that puts out every nth sample of F3
can produce only frequencies that are multiples of 39.14 Hz. Clearly
these offer too limited a choice of frequencies.
The actual algorithm used in 0SC, which overcomes these limitations,
is
S1+ 1 = Sl + 121
0 1 = Ill' F n([Sl]MOd 511)
where
i is the index of acoustic output samples;
Sl is a running sum which increases by 121 for each successive value
of i; Sl is usually set to zero at the beginning of each note;
[StlMOd 511 is [Sl - n· 511] where n is selected so that [SdMOd 511
always falls between 0 and 511;
IIi is the amplitude input that multiplies the amplitude of Fn;
12i is the frequency controlling input; and
0 1 is the output.
The operation of 0SC can be understood geometrically by referring
to Fig. 30. Sl is a ramp function whose slope is 12 units per sample of
acoustic output. [Sl]MOd 511 is the sawtooth function which is reset to
zero each time Sl equals a multiple of 511. With a slope of 12, exactly
511/12 samples are required for Sl to reach 511; hence the period of
[Si]MOd 511 is exactly 511/12 samples. At a sampling rate of 20,000 Hz,
the frequency of [Sl]MOd 511 is
F _ 20,000·12
req - 511
52 CHAPTER TWO

\.----,,----' Samples of acoustic output

IDl
12

Fig. 30. Operation of 0SC.

This is the fundamental relation between the frequency of 0SC and 12.
I t can be written

Freq = 39.4·12

or in case we want to solve for 12 for a given frequency

12 = .02555· freq

M ore generally,

NF' f req
12 = -
R

where NF is the length of a stored function (NF = number of

samples - 1) and R is the acoustic sampling rate.
[Si]Mod 511 has the desired frequency but the wrong waveform-a
simple triangle. [StlMod is used to scan F n as specified by the second
0SC equation. The scanning process is equivalent to projecting samples
of [Si]MOd 511 to the left in Fig. 30 and sa~pling Fn as indicated. This
process, along with a multiplication by 11, gives an output of the desired
frequency, amplitude, and waveform.
Although [StlMod 511 lies between 0 and 511, it will not, in general,
take integer values. Since Fn(k) is sampled and stored only for integer
values of k, some accommodation must be made. The simplest 0SC
algorithm truncates [SdMOd 511 to the next smaller integer value. More
complex 0SC routines interpolate Fn(k) between successive k's.
 A SEQUENCE OF TUTORIAL EXAMPLES 53
In the score 0SC would be called by a statement such as
0SC P5 P6 B2 F2 P30 ;
where P5 is the amplitude input, P6 the frequency input, B2 the 1-0
block for output, F2 the stored function, and P30 is a vacant-note
parameter location for storing the sum S1. One of the note record
parameters must be reserved for S1. Since the initial value of S1 is zero,
the parameter need not be written; unwritten parameters are always set
to zero at the beginning of each note.

Examples of Simple Instruments

Having now discussed the four simplest and most important
generators, let us look at some examples of instruments constructed of
these generators. For each instrument we will show the score cards
which define the instrument and play a note or two. The instruments
will require two or more stored functions. Although the GEN score
cards that generate these functions are shown here we will postpone
until later a detailed discussion of the GEN routines.
We will also postpone discussion of the conversion function which
greatly simplifies writing scores of the notes. Consequently, our scores
will be somewhat labored and should not be considered typical.

Instrument with Attack and Decay

The simplest instrument shown in Fig. 27 produces sounds by turning
an 0SC on and off suddenly. The sudden transients might be heard as
unwanted clicks. An instrument is shown in Fig. 31a with an envelope
that gradually increases the sound amplitude at the beginning of the
note and decreases the amplitude at the end.
The upper 0SC generates the desired envelope which forms the
amplitude input for the lower 0SC. Fl, the waveform function for the
upper 0SC, is the desired envelope as sketched in Fig. 31 b. The 0SC
is used in a degenerate mode in that its frequency will be set at the value
that permits it to go through exactly one cycle of oscillation during the
note being played. Usually this is a very low frequency; however, unlike
real oscillators, computer-simulated oscillators can produce low
frequencies with ease and precision. The frequency-control equation for
0SC is

511
P6 = 20,000· freq
54 CHAPTER TWO

FI

P5 P6 P7
I~ 1 I

,
o 20 491 511

(b)

J=60
F2

1 J
(d)
0

(0 )
-I

(c)

1 INS 0 1 ;
2 0SC P5 P6 B2 Fl P30 ;
3 0SC B2 P7 B2 F2 P29 ;
4 0UT B2 Bl ;
5 END;
6 GEN 0 1 1 0 0 .99 20 .99 491 0 511
7 GEN 0 1 2 0 0 .99 50 .99 205 -.99 306 -.99 461 0 511
8 N0T 0 1 2 1000 .0128 6.70 ;
9 N0T 2 1 1 1000 .0256 8.44 ;
10 TER 3 ;

(e)

FI FI

(f) (g)

Fig. 31. Instrument with attack and decay: (a) block diagram; (b) envelope
function; (c) waveform function; (d) conventional score; (e) computer
score; (f) pianolike envelope; (g) brasslike envelope.
 A SEQUENCE OF TUTORIAL EXAMPLES 55

If we wish exactly one cycle of oscillation per note,

1
Freq = note duration
or
P6 = 511
20,000 x note duration
.02555
note duration
Thus, for the first note, whose duration is 2 sec, P6 equals .0128 (line 8
of score) and for the second note, whose duration is 1 sec, P6 equals
.0256.
The envelope does much more than eliminate clicks. It is as important
in the determination of timbre as the waveform. The attack time is
especially important; percussive instruments have very short times
(1 or 2 msec), stringed instruments having long times (50-200 msec).
In addition, envelopes can have other shapes: the triangular shape shown
as an alternate envelope on Fig. 31 f is typical of a piano, and the envelope
with initial overshoot in Fig. 31g is typical of a brass instrument.
The score, Fig. 31e, is similar to the score in Fig. 27. A few points
should be mentioned. The instrument is named "1 " and is referred to
as "1" in P3 of the N0T cards. 1-0 block B2 is used for both the input
and output of the lower 0SC. This is permissible since all the unit
generators read their inputs before storing their outputs. However, as
will be pointed out later, an 1-0 block must not be used for two
different purposes at the same time. The upper 0SC uses P30 to store its
Si; the lower 0SC uses P29. In general, since the S/s of different
0SC's are different, they must be kept in different locations.
Record 6 causes the generation of the envelope function by evoking
G ENI ; its operation is the same as in Fig. 27. The envelopes produced
by an 0SC have the unfortunate characteristic that the whole envelope
stretches and shrinks with the duration of the note. Thus the attack
time and the decay time are proportional to duration; the second note
in the score will have half the attack time of the first note. Usually this
variation is undesirable since it changes the timbre of the note. Special
attack and decay generators, which avoid this problem, will be taken
up later.
Adding Vibrato
Vibrato, which we will define as a variation in pitch, adds much
interest to tone color. In Fig. 32a 0SC #2 and AD2 have been appended
to the simple attack and decay instrument to provide vibrato. They
 P5 P6 P7 P8 P9

(0)

F3 F3

, ,0
0 511

J = 60 J = 60

j
1 J .,. .,. Glis
(b) (c)

1 INS 0 2 ;
2 0SC P5 P6 B2 Fl P30 ;
3 0SC P8 P9 B3 F3 P29 ;
4 AD2 P7 B3 B3 ;
5 0SC B2 B3 B2 F2 P28 ;
6 0UT B2 Bl ;
7 END;
8 GEN 0 1 1 0 0 .99 20 .99 491 0 511 ;
9 GEN 0 1 200 .99 50 .99 205 -.99 306 -.99 461 0 511 ;
10 GEN 0 2 3 1 1 ;
11 N0T 0 2 2 1000 .0128 6.70 .067 .205 ;
12 N0T 2 2 1 1000 .0256 8.44 .084 .205 ;
13 TER 3 ;
10' GEN 0 1 3 0 0 .999 511 ;
11' N0T 0 2 2 1000 .0128 6.70 4.55 .0128 ;
12' N0T 2 2 1 1000 .0256 11.25 0 .0256 ;

(d)
Fig. 32. Instruments with vibrato or glissando: (a) block diagram; (b) F3
and score for vibrato; (c) F3 and score for glissando; (d) computer score.

56
 A SEQUENCE OF TUTORIAL EXAMPLES 57

provide a time-varying frequency control to 0SC #3, thus producing a

frequency variation in its output. This illustrates that the frequency
control of an 0SC does not have to remain constant over a note, but can
change in any desired way. P7 controls the average pitch. PS determines
the maximum variation in pitch. P9 determines the rate of variations,
which for typical instruments might be 4 to 8 changes per second.
The wave shape F3 of 0SC #2 determines the way in which frequency
changes with time. The exact shape is usually not critical and a sine
wave, as shown, is usually satisfactory.
In the first score card, Fig. 32d, the instrument is named "2" and
referred to as such in the P3 fields of the N0T records. An additional
I -0 block B3 is required by the instrument. Block B2 must hold the
output of 0SC #1 until 0SC #3 has used it as amplitude input. Conse-
quently, 0SC #2 and AD2 has to use B3 to hold the frequency input for
0SC #3. However, after 0SC #3 has completed its computation, both
B2 and B3 are available for other uses; in this case B2 was used to hold
the output of 0SC #3.
The order of computation is the order in which generators are
written in the score. It is essential to maintain the right order. In the
example, 0SC #1 must be written ahead of 0SC #3 since it provides an
input to 0SC #3. 0SC #2 must be written ahead of AD2, and AD2 must
be ahead of 0SC #3 for the same reason. 0SC #1 could be in any order
with respect to 0SC #2 and AD2.
The two GEN1 functions (records Sand 9) are the same as before.
Record 10 calls upon GEN2 to provide a sine wave for F3. P2 = 0 says
to compute F3 at t = 0 with respect to the acoustic output. P3 = 2 says
to call upon GEN2; P4 = 3 says to compute F3; P5 = 1 says to com-
pute the fundamental with amplitude of 1; P6 = 1 says that there is
only one harmonic (i.e., the fundamental).
In note records 11 and 12, P7, PS, and P9 concern pitch and hence are
of special interest. The rest of the parameters are the same as in Fig.
31. P7 determines average pitch. Thus for the first note C 262
P7 = 262 x .02555 = 6.70
PS is set equal to 1% of P7 so that the maximum frequency deviation
will be 1% of the center frequency. Thus the frequency will change from
259.4 Hz to 264.6 Hz. A 1% vibrato is quite large; i% is more typical
of actual players. However, there is much individual variation in
vibrato. P9 determines the number of complete cycles of change per
second, which we have set at S. Thus
P9 = S x .02555 = .205
58 CHAPTER TWO

With a change in F3, and the meaning of P7, P8, and P9, the same
instrument can also be used for glissando. An F3 consisting of a
straight" interpolating" function appropriate for glissando is shown in
Fig. 32c. P9 now becomes

P9 = .0255
duration of note
and causes 0SC #2 to produce one cycle per note (the same as 0SC #1).
P7 is set at
P7 = .0255 x initial note frequency
and P8 at
P8 = .0255 x (final note frequency - initial note frequency)
The action of AD2 and 0SC #2 with F3 is such that at the beginning of
the note B3 will contain .0255 x initial note frequency, and at the end
of the note it will contain .0255 x final note frequency.
Substitution of cards 10', 11', and 12' into the score in place of
cards 10, 11, and 12 will produce the glissando sample shown. Note
that for the second note (A 44o), which has a constant frequency, P8
is equal to 0 since the initial and final frequencies are the same. P6 and
P9 have the same values, and hence P9 could be eliminated if the
instrument were redefined.
The glissando obtained in this way has a linear change of frequency in
hertz. This means that the musical intervals will change faster at the
beginning of the slide than at the end. Although a linear change of
musical intervals might be preferable, this glissando has been much
used and seems perfectly satisfactory. During most slides, listeners are
insensitive to the precise time course of the pitch.

Instrument with Swell and Diminuendo

In the glissando instrument, 0SC #2 and AD2 form a linear inter-
polating unit which generates a frequency control that goes from initial
to final frequency. If we apply the interpolating unit to the amplitude
control on an 0SC, we can obtain a continuously changing amplitude
for crescendos and decrescendos. An instrument with this feature is
shown in Fig. 33.
In order to simplify the score, we have complicated the interpolater
with an extra oscillator 0SC #2. The glissando instrument req.uired
writing the initial frequency in P7 and the (final-initial) frequency in
P8. The swelling instrument is arranged so the initial amplitude is
 A SEQUENCE OF TUTORIAL EXAMPLES 59

P5 P7 P6 P7 P8

#1 '~
o 511

'k:::J
o 511

1 INS 0 3 ;
2 0SC P5 P7 B2 F3 P30 ;
3 0SC P6 P7 B3 F4 P29 ;
4 AD2 B2 B3 B2 ;
5 0SC B2 P8 B2 F2 VI ;
6 0UT B2 Bl ;
7 END;
8 GEN 0 1 3 .999 0 0 511 ;
9 GEN 0 1 4 0 0 .999 511 ;
10 GEN 0 1 200 .99 50 .99 205 -.99 306 -.99 461 0 511 ;
11 N0T 0 3 2 0 2000 .0128 6.70 ;
12 N0T 2 3 1 2000 0 .0256 6.70 ;
13 TER 3 ;

Fig. 33. Instrument with swell and diminuendo.

written in P5 and the final amplitude in P6. 0SC #1 and 0SC #2 both
generate one cycle per note of waveforms F3 and F4, respectively. F3
goes linearly from 1 to 0 over the course of a note and is multiplied by
the initial amplitude in 0SC #1. Similarly, F4 goes from 0 to 1 and is
multiplied by the final amplitude. Thus the output of AD2 will proceed
linearly from the initial amplitude to the final amplitude.
Records 11 and 12 in the score play what amounts to a single note
made up of two notes tied together. The first note swells from 0 to
maximum amplitude, the second decays back to zero. Amplitude
controls in P5 and P6 are obvious. P7 is set to produce one cycle per
note in both 0SC #1 and 0SC #2.
60 CHAPTER TWO

One peculiarity is introduced by the structure and use of the instru-

ment. We want the two notes to blend into each other with no break
between notes. To achieve this, we have omitted the usual attack and
decay 0SC. However, the waveform 0SC #3 must also produce a
continuous output over the juncture. If we were to store the sum for the
0SC in an unused note parameter (P30, for example), it would be reset
to zero at the beginning of each note, a sudden change of phase between
notes would result, and a click might be introduced. To avoid this
transient, the sum is stored in variable VI. The training orchestra
provides space for 200 variables, denoted VI through V200. These
variables may be changed by either the instruments or the score, but
they are not reset at the beginning of a note. Consequently, storing the
sum of 0SC #3 in VI assures that it will never be reset and that the
oscillator will proceed continuously between all notes. However, this
instrument will be limited to playing only one voice.
There are many other uses for variables, as we will see in the next
example.

Instrument that Varies Waveform with Amplitude

We conclude these examples of simple instruments with a not-so-
simple one. It has been shown that one of the factors that contribute
interest to the timbre of real instruments is a change in spectrum with the
intensity of the sound. Usually the loud sounds have more high-
frequency components than the soft sounds. Figure 34 shows an
instrument that is able to change spectrum with amplitude.
The instrument is an elaboration of the swell and diminuendo
instrument shown in Fig. 33, and it uses the same parameters on the
note records. 0SC #1, 0SC #2, and AD2 #1 form a linear interpolation
unit with P5 as the initial amplitude and P6 the final amplitude. These
inputs range from 0 to 1 with 1 as the maximum output. We will call the
instantaneous amplitude Ampl' Ampl is stored in block B2. MLT #1
and AD2 compute B3 according to the relation
B3 = 1 - Ampl = Amp2
Oscillator 0SC #4 is controlled by Ampl' and 0SC #5 by Amp2' Thus
when Ampl is 0, Amp2 is equal to 1, and all the output comes from
0SC #5; when Ampl is 1, Amp2 is equal to 0, and all output comes
from 0SC #4. At intermediate values of Ampb intermediate portions of
output come from 0SC #5 and 0SC #4. In this way the waveform of
F2 in 0SC #5 controls the spectrum at low amplitudes, and the wave-
form of FI in 0SC #6 controls at high amplitudes.
 A SEQUENCE OF TUTORIAL EXAMPLES 61

P5 P7 P6 P7 P8 V3(01) V4{,2)

#1

82

#4
FI

Fb-
F2

1 INS 0 4 ;
2 0SC P5 P7 B2 F3 P30 ;
3 0SC P6 P7 B3 F4 P29 ;
4 AD2 B2 B3 B2 ;
5 MLT B2 VI B3 ;
6 AD2 B3 V2 B3 ;
7 MLT P8 V3 B4 ;
8 0SC B4 V4 B4 F5 P28 ;
9 AD2 P8 B4 B4 ;
10 AD2 B4 V5 B5 ;
11 0SC B3 B5 B5 F2 V7 ;
12 0SC B2 B4 B4,Fl V8 ;
13 MLT B2 B4 B4 ;
14 MLT B4 V6 B4 ;
15 0UT B4 Bl ;
16 END;

Fig. 34. Instrument that varies waveform with amplitude.

62 CHAPTER TWO

The amplitude of the sum of 0SC #4 and 0SC #5 is relatively

independent of Ampl' The normal dependence is restored by MLT #3.
The output of MLT #3 ranges from -1 to + 1; MLT #4 increases this
range to -2047 to +2047, the normal amplitude range.
Frequency control of 0SC #4 and 0SC #5 is a vibrato circuit plus
AD2 #4, which makes 0SC #5 4 Hz higher in frequency than 0SC #4.
A slight divergence adds richness to the tone. The amplitude of the
vibrato is automatically set at 1% of the center frequency of the tone
by MLT #2. This is an expensive way of controlling amplitude, and
better ways will be discussed when C0NVT functions are considered.
The frequency of vibrato is set at about 6 Hz by V4.
The instrument requires six constants as inputs. These are stored in
VI through V6: VI = -1, V2 = 1, V3 = .01, V4 = .2, V5 = .105,
and V6 = 2047. The record that stores these constants is
SV3 0 1 -1 1 .01 .2 .105 2047;
PI and P2 say to set variables in Pass III at time O. P3 says to start with
variable 1 and continue with 2, 3, etc., to the end of the data. P4-P9
give the six numbers to be set in VI-V6. New variables can be set at
any time, as previously set variables can be changed, with other SV3
cards. Times of settings and changes are all controlled by P2.
We will not write a score for this instrument since, except for setting
variables, little new is involved. A reasonable choice for Fl and F2 is
sketched in Fig. 34. The harmonics of Fl decrease at about 6 dB per
octave; those of F2 at 12 dB per octave. Thus the instrument is likely to
have higher-frequency energy at high output amplitudes. Other more
interesting examples of F 1 and F2 could be devised.

C0NVT Function to Process Note Parameters

Scores for the instruments thus far discussed contain many affronts
to a lazy composer, and all composers should be as lazy as possible
when writing scores. For example, computing the frequency control of
an oscillator as
12 = .02555 x frequency in hertz
is a tedious process. Instead, one would like to write the notes of a
scale directly, such as the numbers 0-11 for a 12-tone scale.
A F0RTRAN routine named C0NVT is called at the end of Pass II;
it can apply the full power of F0RTRAN to convert the note parameters
as written by the composer into a new set of parameters, which are the
 A SEQUENCE OF TUTORIAL EXAMPLES 63

inputs to the instruments. As will be clear from the examples below,

the nature of C0NVT depends on the instruments used with it. Conse-
quently, no universal C0NVT program is or can be supplied with
Music V; instead the composer must write his own for each orchestra
he defines. Let us explore the possibilities with the simple attack and
decay instrument designed in Fig. 31.
We shall assume that the composer would like to write frequency
directly in hertz and would like to write amplitude on a decibel scale
rather than on a linear scale. Furthermore, the note duration is already
written in P4; it is an indignity to have to write P6 (= .02555/duration).
Hence we will assume that the composer will write
P5 = amplitude of note in decibels with 66 dB corresponding to a
maximum amplitude of 2000
P6 = frequency of note in hertz
With these inputs C0NVT must compute 2
P5 = 10.0**(P5/20.0)
P7 = 511.0 * P6/(sampling rate)
and
P6 = 511.0/CP4 * sampling rate)
A program to achieve these conversions is given below along with
annotated comments.
Text Notes

SUBR0UTINE C0NVT
C0MM0N IP, P, G
DIMENSI0N IP(10), P(100), G(1000)
IF (P(1) - 1.0) 102, 100, 102 2
100 IF (P(3) - 1.0) 102, 101, 102 3
101 P(5) = 10.0 ** (P(5)/20.0) 4
P(7) = 511.0 * P(6)/G(4)
P(6) = 511.0/(P(4) * G(4»
IP(1) = 7 5
102 RETURN
END

Notes
1. The data-record parameters PI-PI00 have been placed by Pass II
in P(1)-P(I00). The IP array contains some pertinent fixed-point
2 Equations relating to programs will usually be written in a F0RTRAN-like
notation.
64 CHAPTER TWO

constants; in particular, IP(1) = number of parameters in the data

record. G is a general memory array for Pass II.
2. This statement checks to see whether the data record pertains to a
note (rather than a GEN or something else). The numerical
equivalent of N0T is 1. Chapter 3, Section 3 lists the numerical
equivalent of all the operation codes.
3. This statement checks to see whether instrument #1 is referred to by
the N0T record. Other instruments would usually require other
C0NVT functions.
4. These statements perform the desired conversions. The sampling
rate is always kept in variable G(4). Thus in calculating P(6) and
P(7) we have divided by G(4) rather than by the number 20,000.
This is desirable because sampling rate is often changed and, if
C0NVT always refers to G(4) to obtain the current rate, it will not
have to be reassembled with each change of rate. Instead only G(4)
need be modified, and this is a simple change which we will discuss
shortly.
5. C0NVT has added one parameter P(7); thus the word count IP(1)
must be changed to 7. The possibility of generating additional
parameters with C@NVT is most important and attractive since the
composer does not have to write these parameters. In addition, Pass
I and Pass II do not have to process and sort these additions, which
increases efficiency.

With this C0NVT function the score lines to play the two notes on
Fig. 3ld (equivalent to lines 8 and 9 on Fig. 3lc) are
N0T 0 1 2 60 262 ;
N0T 2 I 1 60 330 ;

Now let us construct a somewhat more complicated C0NVT func-

tion for instrument 2 in Fig. 32. We will again use P5 as amplitude in
decibels. Frequency will be specified in terms of an octave, and a
l2-tone note within the octave by P6 and P7, P6 giving the octave and
P7 the step within the octave. Thus, for example

Note P6 P7

-1 o
-1 1
-1 2
 A SEQUENCE OF TUTORIAL EXAMPLES 65

-1 11
o o
o 1

o
The vibrato controls will be eliminated from the N0T record. In-
stead, we will assign two Pass II variables, G(50) to control the percent
frequency variation and G(51) the rate of vibrato.
The equations which must be programmed into C0NVT are
Frequency = 262.0 * (2 ** (P6 + P7/12.O))
P5 = 10.0 ** (P5/20.0)
P6 = 511.0/(P4 * sampling rate)
P7 = 511.0 * frequency/sampling rate
P8 = 511.0 * frequency * G(50)/(sampling rate * 100)
P9 = 511.0 * G(51)/sampling rate
Most of the equations are self-explanatory. The note frequency is com-
puted in hertz from the logarithmic scales embodied in P6 and P7 by
the first relation. The factor 100 is put in the denominator of P8
because G(50) is a percentage.
Vibrato control is a good example of the use of Pass II memory in a
composition. Except for the first few variables, numbers in the G array
may be used for any purpose desired by the composer. Numbers are
placed in the array by an SV2 record, which is analogous to the SV3
record that was previously used to set a Pass III variable. Thus
SV2 0 50 .5 6 ;
would set G(50) = .5 and G(51) = 6 at t = O.
The program to carry out the computations follows.
Text Notes
SUBR0UTINE C0NVT
C0MM0N IP, P, G
DIMENSI0N IP(10), P(100), G(1000)
IF (P(1) - 1.0) 102, 100, 102
100 IF (P(3) - 2.0) 102, 101, 102
101 P(5) = 10.0 ** (P(5)j20.0)
P(7) = 511.0 * 262.0 * (2.0 ** (P(6) + P(7)j12.0))jG(4)
P(6) = 511.0j(P(4) * G(4))
P(8) = P(7) * G(50)j100.0 2
P(9) = G(51) * 511.0jG(4) 3
IP(2) = 9
102 RETURN
END
66 CHAPTER TWO

Notes
1. This statement calculates frequency, multiples it by the appropriate
constant of proportionality, and stores it in P(7).
2. This statement computes the maximum vibrato deviation. The
properly scaled frequency is already available in P(7) and hence must
only be multiplied by G(50)/100.0.
3. This statement sets the rate of vibrato. The constant of propor-
tionality, 51 1.0/G(4) , is the same as any other 0SC frequency
control. G( 51) will be the vibrato frequency in hertz.
A score for this instrument to replace lines 11 and 12 on Fig. 32 is
SV2 0 50 1 6;
N0T 0 2 2 60 0 0 ;
N0T 2 2 1 60 0 4 ;
Once G(50) and G(5I) are set, any number of notes may be written
with the same vibrato. On the other hand, the vibrato constants may be
changed at any time by a subsequent SV2 record. For example, the
deviation could be reduced and the rate increased at t = 15 sec by the
record
SV2 15 50 .5 8 ;
One may ask, why use Pass II variables in C0NVT rather than Pass
III variables or Pass I variables which also exist? The answer is,
C0NVT is a Pass II subroutine and can only make use of information
available in Pass II.
A final example of a C0NVT subroutine will provide a convenient
score language for the glissando instrument in Fig. 32. As shown, the
initial frequency of a note must be written in P7 and the (final - initial)
frequency in P8. We shall eliminate the arithmetic to calculate
(final - initial). Instead the score card will have
P5 = amplitude in decibels
and
P6 = final frequency in hertz
The initial frequency of each note will be defined as the final frequency
of the preceding note. The initial frequency of the first note will be read
into the program with an SV2 card into G(50).
In order to use this simple form the program must remember the
final frequency of each note. G(50) will also be used for this purpose.
The program to achieve these objectives follows.
 A SEQUENCE OF TUTORIAL EXAMPLES 67

Text Notes

SUBR0UTINE C0NVT
C0MM0N IP, P, G
DIMENSION IP (10), P(100), G(1000)
IF (P(1) - 1.0) 102, 100, 102
100 IF (P(3) - 2.0) 102, 101, 102
101 P(5) = 10.0 ** (P(5)j20.0)
P(7) = G(50) * 511.0jG(4) 1
P(8) = (P(6) - G(50» * 511.0jG(4) 2
G(50) = P(6) 3
P(6) = 511.0/(P(4) * G(4» 4
P(9) = P(6)
IP(2) = 9
102 RETURN
END

Notes
1. This statement sets the initial frequency, which was stored in
G(50).
2. This statement computes the (final - initial) frequency.
3. This statement stores the final frequency in G(50) to become the
initial frequency of the next note.
4. This and the following statement set the frequency inputs of 0SC #1
and #2 to 1 cycle per note.
Figure 35 shows a brief score for the instrument. Only the N0T
cards and the SV2 cards are shown. Record 1 sets the initial frequency

~
/ /
/
l, ....
JI ,
"
/
Glis Glis

1 SV2 0 50 262 ;
2 N0T 0 2 2 60 440 ;
3 N0T 2 2 1 60 330 ;
4 N0T 3 2 1 60 330 ;
5 SV2 4.5 50 440 ;
6 N0T 5 2 1 60 400 ;
7 N0T 6 2 2 60 262 ;

Fig. 35. Score for glissando instrument using a C0NVT subroutine.

68 CHAPTER TWO

to 262 Hz. Records 2,3, and 4 generate an initial glissando. An arbitrary

choice is made. The frequency slide is completed at t = 3, and the third
note is held at E 330. Record 5 resets the initial frequency to 440 for the
beginning of the second glissando. The change is done at t = 4.5 in the
middle of a rest and hence is inaudible. A different arbitrary choice for
glissando is made for the last two notes. During the first note
(5 < t < 6), the frequency changes from 440 to 400. During the second
note (6 < t < 8) the frequency shift continues from 400 to 262.
The flexibility of one method for obtaining frequency slides has been
demonstrated. The use of Pass II variables as a memory for the C0NVT
subroutine is important. Powerful logic may be programmed in this
way.

Additional Unit Generators-RAN, ENV, FLT

Next to be discussed are the generator of random signals RAN, the
band-pass filter FLT, and the envelope generator ENV. Although a
few other generators exist and one can easily design his own generators,
the three described here plus the stereo output box are sufficient for
most purposes.

Random Signal Gen era tor-RAN

RAN is a source of random signals with controlled amplitude and
spectrum. The spectrum is low pass and contains energy from zero
frequency to a cutoff frequency determined by one of the inputs (12).
(As will be shown later, RAN may be combined with an oscillator to
obtain a band-pass signal.)
The diagram of RAN is shown in Fig. 36a. 11 controls the amplitude
of 0; 12 controls its bandwidth. The equation of operation is
0 i = IIi· R i (l21)
where R i(l21) is a low-pass random function whose amplitude varies
from - 1 to + 1 and whose cutoff frequency is approximately
(R·I2)/I024 where R is the sampling rate. The amplitude of 0 1 varies
from - IIi to IIi. The approximate spectrum of 0 is shown in Fig. 36b.
The cutoff frequency is not abrupt, and there are lobes of energy above
(R·I2)/I024 Hz. Neither does the main passband have a flat top. Even
with these deficiencies, the generator is very useful.
The key to RAN is the generation of the random function R i . This is
obtained by generating a sequence of independent random numbers Nh
which are uniformly distributed over the interval - 1 to + 1. These
 A SEQUENCE OF TUTORIAL EXAMPLES 69

(O)g~~
9
o
0 t R°r2
1024 Hz
F,,,,,,,,,
(b)
o

(el

Fig. 36. Random generator-RAN: (a) block diagram; (b) spectrum of 0;

(c) medium-frequency random function; (d) high-frequency random
function; (e) low-frequency random function.

numbers are connected by straight lines to form the continuous func-

tions shown in Fig. 36c. R1 consists of samples of the line function, the
sampling interval being chosen so that 512/121 sampling intervals fall
between each N1 (512/121 does not need to be an integer). The algorithm
for sampling the line function is similar to the sum in 0SC.
A medium-frequency function with 12 ~ 128 is shown in Fig. 36c.
It wiggles at a moderate rate and at R = 20,000 Hz would have a
cutoff of about 2500 Hz.
A high-frequency function with 12 ~ 256 and a cutoff of 5000 Hz is
shown in Fig. 36d. The N1 independent random numbers occur more
frequently here, giving the function a more jagged appearance and a
higher-frequency spectrum. The maximum useful value of 12 is 512.
 70 CHAPTER TWO

This will produce an independent number for each sample and will
achieve a cutoff of Rj2, which is the highest frequency representable by
R samples per second.
A low-frequency function with I2i ~ 64 and a cutoff of 1250 Hz is
shown in Fig. 36e. It is clearly the smoothest of the three functions.
A score record to evoke RAN is
RAN P5 P6 B2 P30 P29 P28 ;
where P5 is the 11 input, P6 the 12 input, B2 the output; P30 is an unused
note parameter for the storage of the sum; and P29 and P28 are two
other temporary storage locations.

P5 PEi P8 P7 P5 PEi P7 P8 P9 P5 PEi P7 P8 VI P9 V2

82 82

FI

P
F~
(a) (b) ( c)

(a) INS 0 I (b) INS 0 2 ; (c) INS 0 3 ;

0SC P5 P6 B2 0SC P5 P6 B2 0SC P5 P6 B2
FI P30 ; FI P30 ; FI P30 ;
RAN B2 P8 B2 RAN P8 P9 B3 0SC P8 VI B3
P29 P27 P26 ; P29 P27 P26 ; F3 P29 ;
0SC B2 P7 B2 AD2 P7 B3 B3 ; RAN P9 V2 B4
F2 P28 ; 0SC B2 B3 B2 P28 P26 P25 ;
0UT B2 BI ; F2 P28 ; AD3 P7 B3 B4 B3 ;
END; 0UT B2 BI 0SC B2 B3 B2
END; F2 P27 ;
0UT B2 BI ;
END;

Fig. 37. Examples of instruments with random generator: (a) amplitude-

modulated band-pass noise; (b) frequency-modulated band-pass noise;
(c) periodic plus random vibrato.
 A SEQUENCE OF TUTORIAL EXAMPLES 71

Three useful instruments involving RAN are shown in Fig. 37.

Instrument 1 produces a band-pass noise by amplitude modulation of an
0SC with RAN. Both the center frequency and the width of the pass-
band are controlled by note parameters, P7 determining the center
frequency and P8 the width.
The top 0SC produces the initial attack and decay with function Fl.
The bottom 0SC has a sinusoidal waveform F2, and without RAN it
would produce a single-frequency sinusoid at R·P7/511 Hz. By virtue
of the multiplication inherent in the amplitude input to 0SC, the
sinusoid is multiplied by the output of RAN. Thus the output of RAN
is modulated by the sinusoid and according to the convolution theorem
(Appendix B) the band-pass spectrum sketched in Fig. 38 is achieved.
Spectrum of RAN output

_......::::::IL...-I.~----rz::::::o..-..... Frequency
1~24' P8 Hz
5~2' P8 Hz

Frequency

Fig. 38. Spectrum of instrument that generates amplitude-modulated

band-pass noise.

The modulation can be looked upon as shifting the (low-pass) spectrum

of RAN and centering it about the frequency of 0SC.
If F2 is a complex wave with harmonics, the modulation will generate
a replica of the RAN spectrum centered about each harmonic, or a
multiple-band noise. Because the auditory effect is usually muddy and
unpleasant, most instruments use a sinusoidal F2.
The center frequency of the passband is simply the frequency of
0SC-R·P7j511. The bandwidth is two times the cutoff frequency of
RAN or R·P8j512. The factor 2 comes from the negative frequencies
in the RAN spectrum which are shifted into positive frequencies by the
modulation. The C0NVT function appropriate to this operation is
511
P7 = If . center frequency

and
72 CHAPTER TWO

P8 = 512. bandwidth
R

It is often desirable to make the bandwidth a fixed percentage of the

center frequency. This corresponds to a fixed musical interval about the
center frequency. The C0NVT equation is simply
P8 = k·P7
and P8 must not be written in the score.
Very narrow bands of noise can be generated by small values of P8.
In fact, for P8 = 0, a zero bandwidth or pure sinusoid is produced.
Narrow-band noises produced by amplitude modulation reveal the
way in which they are generated; they sound like a sinusoid with a
fluctuating amplitude. This sound is often not what the composer
desires or expects from such noises; however, it is an essential
characteristic of amplitude modulation and cannot be avoided with this
instrument.
INS 2 (Fig. 37b) generates frequency-modulated noise with a band-
pass spectrum. The center frequency of the band is again controlled by
P7 as P7· R/511. The rest of the characteristics of the spectrum are not
as easy to estimate as in the case of the amplitude modulated noise.
P8· R/511 is the maximum instantaneous deviation of frequency of
0SC. Frequency-modulation theory says that the width of the noise
band will be somewhat greater than 2· P8 . R/511. For most purposes
2· P8 . R/511 is a useful estimate of bandwidth.
P9 determines the rate at which the frequency of 0SC deviates. Its
effect on the spectrum is hard to compute precisely. Experience has
indicated that in order to produce" smooth" sounding noise, P9 should
be about five times P8. C0NVT is a convenient place to set both P8 and
P9.
At very small bandwidths, INS 2 sounds like a sine wave with a small
random variation in frequency.
INS 3 (Fig. 37c) shows an excellent vibrato circuit devised by J. C.
Tenney. The frequency variation contains a periodic component supplied
by 0SC #2 and a random component supplied by RAN. A useful set of
parameters is

P8 = P9 = .0075 * P7
to give i-percent periodic and i-percent random variation

VI = 8 * 511/R
 A SEQUENCE OF TUTORIAL EXAMPLES 73

for an 8-Hz periodic vibration rate and

V2 = 16 * 511/R
for a 16-Hz random bandwidth. The random bandwidth tends to be
substantially greater than the periodic frequency.

Envelope Generator-ENV
The use of 0SC as an envelope generator is satisfactory in some
applications, but it makes the attack and decay times proportional to
the total note duration. Important aspects of timbre depend on the
absolute attack time. With 0SC, these will change from long notes to
short notes. The difference may be enough to give the impression of
two different kinds of instruments.
A special generator ENV has been programmed to sweep away this
limitation. It allows separate control of attack time, steady-state dura-
tion, and decay time. In order for ENV to be effective, a special
C0NVT function must be written for ENV. The computations in
C0NVT are at least as complex as those in ENV.
An instrument using ENV is illustrated in Fig. 39. ENV has four
inputs II-14 and requires one function. II determines the amplitude of
the output, and 12, 13, and 14 the attack time, the steady-state time, and
the decay time, respectively. The function Fl is divided into four equal
sections, the first determining the shape of the attack, the second the
shape of the steady state, and the third the shape of the decay. The last
section is not used and should be zero to allow for any round-off error
involved in scanning the first three parts.
The output 0 1 may be written
0 1 = IIi * function (scanned according to 11, 12, and 13)
The first quarter of the function is scanned at a rate of 12 locations per
sample, the second quarter at a rate of 13 locations per sample, and the
third quarter at a rate of 14 locations per sample. Consequently,
, C0NVT should compute
12 = 128
attack time * sampling rate
13 = 128
steady-state time * sampling rate
and
14 = 128
decay time * sampling rate
74 CHAPTER TWO

P5 P9 PIO PII P7 P8 VI
F3

II
fv-
Per , j
-e-
J =300
J.
. ~

T J
FI

5/1

1 INS 0 4
2 ENV P5 Fl B2 P9 PlO P11 P30 ;
3 0SC P8 VI B3 F3 P29 ;
4 AD2 P7 B3 B3 ;
5 0SC B2 B3 B2 F2 P28 ;
6 0UT B2 Bl ;
7 END;
8 GEN 0 1 1 0 0 96 1 128 .7 150 1 175 .6 200 1 225 .7 256 1
320 .3 384 0 511 0 ;
9 SV2 0 50 .050 .100 ;
10 SV3 0 1 .15 ;
11 N0T 0 4 .1 54 349 ;
12 N0T.2 4 .1 54 392 ;
13 N0T.5 4 .13 54 440 ;
14 N0T.6 4 .2 54 349 ;
15 N0T 0 4 .8 54 262 ;

Fig. 39. Envelope generator ENV for attack and decay. Instrument #4.
The attack time AT and decay time DT will in general be constants.
C0NVT calculates the steady-state time SS as the duration PC4)
minus the attack and decay times
SS = PC4) - AT - DT
 A SEQUENCE OF TUTORIAL EXAMPLES 75

Thus the steady-state time varies with duration. For short notes there
may be no steady state, and the attack and decay times may have to be
shortened so that their sum does not exceed the duration. All this must
be done by C0NVT.
The data record to evoke ENV is
ENV II, F, 0, 12, 13, 14, S ;
S is a sum that must be assigned temporary storage in some unused note
parameter.
In the example shown in Fig. 39, Pass II variables V50 and V5I
contain the attack and decay times, respectively. These are set with the
SV2 record. The vibrato rate is kept in Pass III VI and is set with SV3
record. The attack and decay function FI is computed with GEN1. The
attack portion has a slight overshoot for added sharpness. The steady-
state portion has two cycles of quaver. The decay portion has two line
segments to approximate an exponential.
The N0T records contain amplitude in decibels in P5 and the
frequency in hertz in P6.
The instrument requires inputs P5 and P7 through PI1, as shown on
the diagram. The C0NVT program to compute these inputs from the
N0T record is listed and annotated below.

Text Notes

SUBR0UTINE C0NVT
C0MM0N IP, P, G
DIMENSI0N IP(10), P(100), G(1000)
IF (P(l) - 1.0) 105, 100, 105
100 IF (P(3) - 4.0) 105, 101, 105
101 COR = 1.0 1
SS = P(2) - G(50) - G(5l)
IF (SS) 102, 103, 103 2
102 C0R = P(4)/(G(50) + G(5l) 3
P(10) = 128.
G0 T0 104
103 P(10) = 128./(G(4) * SS) 4
104 P(9) = 128./(G(4) * G(50) * COR) 5
P(ll) = 128./(G(4) * G(5l) * COR)
P(5) = 10.0 ** (P(5)/20.0) 6
P(7) = 511.0 * P(6)/G(4)
P(8) = .0075 * P(7)
IP(l) = 11
105 RETURN
END
76 CHAPTER TWO

Notes
1. C0R will correct attack and decay times for short notes where the
steady state does not exist.
2. Checks to see if steady state time SS is positive.
3. Steady state is negative. C0R is set to reduce attack and decay times
proportionally so that
AT + DT = duration
P(10) is set at 128 so that steady-state time will equal one sample,
which is the minimum possible steady state.
4. Computation of P(10) for positive steady-state times.
5. Computation of P(9) and P(ll) for either positive or zero steady-
state times. C0R will be less than 1 for zero steady-state times.
6. The usual computation of amplitude and frequency control. Vibrato
amplitude is set at i percent of center frequency.
The N0T records (11-15) play the five notes sketched on the staff.
Two additional capabilities of the program are inherent in these records.
Instrument #4 is used to play up to three voices simultaneously. The
second voice is a sustained C 262 • The third voice occurs because the
slurred notes overlap slightly, with the note from record 6 extending into
the beginning of the note from record 7. Pass III can play multiple
simultaneous voices on any instrument. As many as 30 voices can be
played in the training orchestra.
The score records are not written in ascending sequence of action
times, in that the C 262 is written last and starts at t = 0. The order of
these records is immaterial, since they will be sorted into the proper
ascending sequence of action times in Pass II.
Filter-FLT
One of the more difficult sound-processing operations is filtering. A
unit generator that operates as a band-pass filter is shown in Fig. 40.
The filter may be used to introduce formants or energy peaks at specified
frequencies into sound waves. Such formants are characteristic of many
instruments.
The filter is calculated by means of a difference equation. In terms of
the diagram shown in Fig. 40, the equation is
0 1 = III + 121 ·0i - 1 - 131 ·0i - 2
11 is the input to the filter, 0 the output, and 12 and 13 determine the
frequency and bandwidth of the passband.
More specifically, the difference equation approximates a 2-pole
filter with a pole pair at - D ± j F Hz on the complex frequency plane
 A SEQUENCE OF TUTORIAL EXAMPLES 77

1'1
F Hz

-..j 0

IZ x
FLT

13 F

....!L-Hz
2'1J"

( a)
(b)

Gam

Go~----

O~----------:!F~-------"Freq Hz
(e)

Fig. 40. Band-pass filter-FLT. (a) Diagram: 12 = 2e- 2nD / R cos 21TF/R;
I3 = e- 4nD /R ; (b) poles in complex plane; (c) curve of gain vs. frequency.

as shown in Fig. 40. The approximate gain of the filter is also shown as a
function of frequency in Fig. 40, where the peak occurs at F Hz and the
bandwidth at the half-power points is 2D Hz. The approximation holds
for F» D. 12 and I3 are determined as functions of F and D by the
relations
12 = 2 e- 2nD/R cos 27TF /R
and
I3 = e- 4 :n:D/R
where R is the sampling rate. C0NVT may be conveniently used to
compute 12 and 13 from F and D.
78 CHAPTER TWO

The main problem in using the filter is obtaining a reasonable ampli-

tude of output. The dc gain Go is given by the equation
1
Go = 1 - 12 + 13
and the peak gain is approximately
F
Gp = GO' for F» D
2D
Both gains may be either less than or much greater than unity, depend-
ing on F and D. Narrow bandwidths produce high peak gains.
The amplitude of the output depends both on the amplitude of the
input and on its frequency composition. A sinusoid near frequency F
will be multiplied by G p • A low-frequency sinusoid will be multiplied
by Go. A complex signal must be decomposed into individual har-
monics, the gain for each harmonic computed separately, and the
resulting amplified harmonics reassembled at the output. This process is
usually impractical, and one approximates the gain as something between
Go and G p • Often the approximation is poor and it is necessary to
adjust the amplitudes and recompute the samples to avoid overloading
or underloading the output. For this reason filters should be used
sparingly.
The score record for FLT is
FLT 11, 0, 12, 13, Ph P j
where Pi and P j are two unused note parameters in which 0 i -1 and
0 1 _ 1 are stored.

Composing Subroutines-PLF
Our tutorial discussion of what might be called the basics of sound
generation is now complete. We are ready to take up compositional
subroutines that will permit the generation of note parameters by the
computer. These are some of the most interesting but difficult directions
in which computer sound generation can be developed. Advanced
applications point toward complete pieces composed by computer.
However, long before these goals are achieved, PLF subroutines will be
useful in saving the human composer from much routine work.
So far, for each note to be played, the composer has had to write a
line of score starting with N0T.... PLF subroutines will now be
developed which write these N0T records. Furthermore, one score
record that evokes a PLF subroutine can generate many N0T's.
 A SEQUENCE OF TUTORIAL EXAMPLES 79

Although PLF programs can do other things besides compute N0T

records, these records are of overriding importance and are the reason
for creating PLF's. Moreover, Pass I itself is justified because it serves
to contain the PLF subroutines.
Let us develop an example to demonstrate and teach the possibilities
and practices of PLF programs. The example will allow storing a group
of N notes in Pass I memory. A call to a PLF1, which we shall write,
will insert these N notes anywhere in a composition and modify the
notes by an arbitrary frequency shift, by an arbitrary tempo shift, and
by specifying the instrument on which they will be played.
Figure 41 illustrates the use of PLFI. The first four score records
store the four-note pattern in Pass I variable storage, using variables
10 through 43. Like the other passes, Pass I has a general storage array
D(2000), which contains 2000 locations in the training orchestra. The

, J J JJ
~>
(a)

I I LJ
5 6 8 9
(b)

1 SVI 0 10 0 1 52 0 ;
2 SVI 0 20 1 1 56 .167 ;
3 SVI 0 30 2 1 60 .333 ;
4 SVI 0 40 3 1 56 0 ;
5 PLF 0 1 10 40 0 1 .583 4 ;
6 PLF 0 1 10 40 4 1 .750 4 ;
7 PLF 0 1 10 40 4 1 1.167 5 ;
8 PLF 0 1 10 40 8 .5 .417 4 ;
9 PLF 0 1 10 10 10 2 .583 4 ;
(c)

Fig. 41. PLF note-generating example: (a) pattern; (b) conventional score;
(c) computer score.
80 CHAPTER TWO

PLFI routine will assume that a note is described by four numbers

Starting time in seconds
Duration in seconds
Amplitude in decibels
Frequency in logarithmic units
in four successive D locations. Successive notes in a pattern will be
separated by 10 locations in D, so that the first note goes into D(10),
the second into D(20), etc., as is accomplished by the SVI records.
The logarithmic frequency scale is introduced here. The composer
will write scale numbers which are related to frequency by the equation
S I - I (freqUency in hertZ)
ca e - Og2 262.0
This is probably the most useful scale for compositional algorithms.
Middle C262 is 0, C512 is + 1, etc.; the even-tempered half step is an
increment of /2. Thus
C 262 =0
C# = .083
D = .167
D# = .250
E = .333
F = .417
G = .583
G# = .667
A = .750
A# = .833
B = .917
C 512 = 1.000
The even-tempered standard musical intervals are
half step = /2 = .083
full step = -i = .167
minor third = 1; = .250
major third = t = .333
fourth = !-2 = .417
fifth = 122 = .583
Frequency transposition can be done simply by adding a constant to
the scale steps of a pattern. Multiplication corresponds to increasing or
decreasing the size of the i1ntervals in a pattern. Scales other than
12-tone can be represented with equal facility. The logarithmic scale is
 A SEQUENCE OF TUTORIAL EXAMPLES 81

so powerful and appropriate that we will use it almost exclusively from

here on.
In Fig. 41 score records 5 through 9 evoke the PLF subroutine that
is to be presented. The P fields in these records have the following
significance.

PI calls a PLF subroutine

P2 is not used since action time has no importance in Pass I
P3 identifies the subroutine number (PLFl)
P4 gives the D location of the first note in the pattern
P5 gives the D location of the last note in the pattern
P6 gives the time in seconds at which the pattern should begin
P7 gives the duration scaling of the pattern; .5 = play at double
speed; 2 = play at half speed
P8 gives the logarithmic interval to shift the frequency of the
pattern. For example, P8 = .583 corresponds to shifting the
theme up by a fifth
P9 gives the instrument number on which the pattern should be
played

PI through P3 have the same significance for all PLF routines. The
rest of the P's depend entirely on the particular subroutine to be written.
The conventional score for the notes produced by records 5 through 9
is shown in Fig. 41 with the notes coming from a given record identified.
Record 5 produces the first four notes in which the pattern is shifted up
by a fifth. Records 6 and 7 produce two copies of the pattern playing in
fourths. The upper voice is played on instrument 5 which is assumed to
yield a staccato timbre. RecoId 8 plays the pattern at double speed.
Record 9 plays the first note of the pattern at half speed.
In order to write a PLF program we will have to know something of
the operation of Pass I. It reads the score records in the order in which
they appear in the score. The SVI records cause data to be stored in the
D(2000) memory. A N0T record would simply cause the record to be
sent on to Pass II. This is accomplished by placing the N0T data in the
P(lOO) array and calling a communication routine WRITE 1, which
writes out the P array on a file that will later be read by Pass II. For
bookkeeping purposes, the number of parameters in the record is kept
in another Pass I location IP(l) and is automatically written out by
WRITEl. The function of the PLF routine is to generate N0T records
and to write them out exactly as Pass I would have done with a record
in the score.
82 CHAPTER TWO

How is the PLF routine brought into action? When Pass I reads a
PLF score record it calls a subroutine PLFn, in which n is in P3.
The rest of the data on the score record is in the P array where it can
be used by the subroutine.
The annotated PLFI routine to perform the computations we have
described follows.

Text Notes

SUBR0UTINE PLFI
C0MM0N IP, P, D
DIMENSI0N IP(10), P(100), D(2000)
NS = P(4) 2
NE = P(5)
TS = P(6)
DS = P(7)
FS = P(8)
IP(1) = 6 3
P(1) = 1.0 4
P(3) = P(9) 5
D0 100 I = NS, NE, 10 6
P(2) = TS + DS * D(I) 7
P(4) = DS * D(I + 1) 8
P(5) = D(I + 2) 9
P(6) = (2.0 ** (D(I + 3) + FS» * 262.0 10
CALL WRITE1(10) 11
100 C0NTINUE
RETURN
END

Notes
1. This C0MM0N and DIMENSI0N statement locates the three
essential arrays, IP, P, and D for PLFI. The Pass I definition of
these arrays must agree with the definition in the subroutine.
2. These statements take parameters P4-P8 from the PLF data record
and store them in the PLF subroutine. Since the P(100) array will
be used to output N0T records, the PLF parameters must be
removed from it.
3. The word count of the N0T records is set at 6. We will assume that
we are generating notes for an instrument of a type shown in Fig.
39. The six fields are

PI N0T
P2 Action time in seconds
 A SEQUENCE OF TUTORIAL EXAMPLES 83

P3 Instrument number
P4 Duration in seconds
PS Amplitude in decibels
P6 Note frequency in hertz

4. The programs convert all alphabetical symbols into numerical

equivalents in the initial reading routine in Pass I. All subsequent
processing is done on the numbers. The equivalence of N0T is
1.0, which is set by this statement.
5. This statement sets the instrument number into P(3). P(l) and P(3)
are constant for all the notes in the pattern and hence can be set
once at the beginning.
6. This D0 loop is executed once for each N0T in the pattern. The
storage of the pattern in the D array is inherently defined by the
loop, the first note beginning at D(NS), the last note beginning at
D(NE), the notes being 10 locations apart.
7. This statement computes the starting time of the note as the PLF
time shift TS, plus the duration scale DS, times the starting time
relative to the beginning of the pattern D(I).
8. This statement computes the scaled duration of the note.
9. This statement transfers the amplitude of the note from the pattern
to the P(100) array. No modification of amplitude is necessary.
10. This statement adds the frequency transposition FS to the pattern
frequency D(I + 3), and converts the sum from a logarithmic to a
linear frequency scale in hertz.
11. This statement calls for writing out the completed N0T record.

In this example it is already possible to see many labor-saving

advantages in PLF. Although the pattern is atypically short, 17 notes
are produced by only five data records, far fewer than are needed to
write a separate N0T record for each note.
Next let us discuss a slightly more complicated and considerably more
interesting PLF routine. It takes the product of two themes, in a sense.
Each note in the first theme is replaced by the entire second theme.
The second theme is scaled for duration; its total duration exactly
equals the duration of the note it replaces. The log frequencies of the
second theme are increased by the log frequencies of the replaced note,
so that the second theme is centered about each note of the first theme.
Amplitudes are similarly treated. The process is reminiscent of some
"theme and development" styles. The second theme can be considered
an ornament applied to the first theme.
84 CHAPTER TWO

An example is shown in Fig. 42, where the first theme, the second
theme, and the product are written in musical notation. Typically, theme
2 is short and compact in frequency range. However, this is not a
requirement. We could also form the product
Theme 2 x theme I
Our multiplication algorithm is not commutative, and
Theme 1 x theme 2 =1= theme 2 x theme 1

, ,1.1 I J i ,j
(0 ) ( b)

, JJ £J D In ill }. .f.Jll.8 JrJ iJ I

(c)

1 SVI 0 10 0 2 50 .417 ;
2 SVI 0 20 2 2 53 .750 ;
3 SVI 0 30 4 1.5 56 .583 ;
4 SVI 0 40 5.5 .5 59 .167 ;
5 SVI 0 50 6 2 62 .417 ;
6 SVI 0 60 0 1.08 4 0 ;
7 SVI 0 70 1.5 .375 2 .167 ;
8 SVI 0 80 2 .75 0 - .083 ;
9 SVI 0 90 3 .75 - 2 0;
10 PLF 0 2 10 50 60 90 0 4 ;
(d)

Fig. 42. Multiplication of two themes: (a) theme 1 times (b) theme 2 gives
(c) product via PLF2; (d) computer score.
The score for the example is also shown in Fig. 42. Lines 1-5 define
theme 1, lines 6-9 define theme 2, and line 10 calls PLF2 to generate the
product. The calling sequence is
P4 D location of first note of first theme
P5 D location of last note of first theme
P6 D location of first note of second theme
P7 D location of last note of second theme
P8 Starting time of product theme
P9 Instrument number
 A SEQUENCE OF TUTORIAL EXAMPLES 85

We will assume that the D array arrangement and the instrument are
the same as were used for PLFI.
The annotated F0RTRAN program follows.

Text Notes

SUBR0UTINE PLF2
C0MM0N IP, P, D
DIMENSI0N IP(10), P(100), D(2000)
NBI = P(4)
NEI = P(5)
NB2 = P(6)
NE2 = P(7)
TS = P(8)
IP(1) = 6
P(1) = 1.0
P(3) = P(9)
D0 101 I = NB1, NE1, 10 2
START = TS + D(I) 3
DS = D(I + 1)/(D(NE2) + D(NE2 + 1»
D0 100 J = NB2, NE2, 10 2
P(2) = START + DS * D(J) 4
P(4) = DS * D(J + 1)
P(5) = D(J + 2) + D(I + 2)
P(6) = (2.0 ** (D(J + 3) + D(I + 3)) * 262.0
CALL WRITE1(10)
100 C0NTINUE
101 C0NTINUE
RETURN
END

Notes
1. This group of statements moves the PLF parameters from the P(100)
array into the subroutine and sets the unchanging parts of the N0T
parameters in P(lOO) in preparation for writing N0T records.
2. The program contains two nested D0 loops. The outer loop is
executed once for each note in theme 1, the inner D0 cycles for each
note in theme 2.
3. These two statements compute the starting time shift and the
duration scaling for a repetition of theme 2. START is the beginning
time of a note in theme 1. DS is computed so that the last note in
theme 2 will end at the ending time of the note in theme 1.
4. This and the following three statements compute the starting times,
durations, amplitudes, and frequencies for the notes in theme 2
86 CHAPTER TWO

which are replacing a single note in theme 1. Amplitudes in decibels

and frequencies on a logarithmic scale are simply added. Frequencies
are converted to hertz.

These two examples give only a slight indication of the range of

objectives that may be programmed with PLF routines. The routines
are not limited to generating N0T records. They may also be used to
manipulate the information stored in the D array of Pass I. A powerful
appli.:~ation is a set of PLF routines, each of which effects a different
transform on a set of notes stored in the D array. Since the result of
each transform must be in the same form as its input, several trans-
forms may be successively applied. Finally, a last PLF writes out the
N0T records. A composition using these subroutines would consist of
the description of some thematic material, plus a long sequence of PLF
calls to manipulate this material.
The PLF routines provide one of the most exciting areas for further
development in the entire Music V structure. Not only do they promise
the most interesting possibilities, but they also offer the greatest
challenges to the composer's creativity.

Compositional Functions
The note-generating subroutines that have just been demonstrated
can be greatly strengthened by defining information in certain ways
which we call compositional functions. Compositional functions can be
used to provide a new language to describe sounds, called a graphic
score. Although graphic scores can be used to represent the notes in a
conventional score, the notation is completely different. It is more
powerful in the sense that many sounds that are impossible to notate
conventionally can be readily described by a graphic score. Moreover,
in the synthesis of sounds, the graphic scores can be "read" easily by
note-generating subroutines. This section can only lay the foundation
for graphic scores, but further information is given in the references.
A compositional function is a function defined over an entire section
of a composition. It is used to control some always present parameter
such as loudness or tempo. Compositional functions should not be
confused with the stored function used to describe waveshape or
envelope. The stored functions are generated, stored, and used in
Pass III. Compositional functions are described and used in the first
two passes. Both their mode of description and their use differ from
those of stored functions.
 A SEQUENCE OF TUTORIAL EXAMPLES 87

Metronome Function
Let us start by considering the metronome function which is built
into Pass II and can be evoked if desired. So far, the starting times and
durations of notes have been written in numbers which were interpreted
as seconds. Thus a note
N0T 2 4 I 54 .167 ;
starts at 2 sec from the beginning of the section and lasts for 1 sec.
With a metronome function, P2 and P4 are interpreted in beats; the
note starts at the beginning of the second beat of the section and lasts
for one beat. The relation between beats and seconds is given by the
metronome function, which is in standard metronome marking of beats
per minute. Thus, for example, if the metronome function is 180, the
note would start i sec from the beginning of the section and would last
i sec.
The metronome function need not be constant, but can change
abruptly or gradually during a section to introduce accelerandos or
retards. The operation can be illustrated by an example shown in
Fig. 43.
The conventional score for 14 quarter notes lasting 14 beats is shown
at the top, together with tempo marking. The N0T cards to encode this
are as follows.

N0T o 4 .8 60 o
N0T 1 4 .8 60 .167
N0T 2 4 .8 60 .333
N0T 3 4 .8 60 o
N0T 4 4 .8 60 o
N0T 5 4 .8 60 .167
N0T 6 4 .8 60 .333
N0T 7 4 .8 60 .417
N0T 8 4 .8 60 .583
N0T 9 4 .8 60 .750
N0T 10 4 .8 60 .917
N0T 11 4 .8 60 .583
N0T 12 4 .8 60 .583
N0T 13 4 .8 60 .750

The metronome function is evoked and is stored in Pass II variable

storage by the records
SV2 0 50 0 60 4 60 8 120 11.9 120 12 30 14 30
SV2 0 2 50 ;
88 CHAPTER TWO

J=60 Accelerando J= 120 J=30

11 J J111J J J 1J J JJ I) j
o 4 8 12
(a)

r:[~::~:':'h o 2 4 6
Beats
8 10 12 14

(b)

<II

'c
:::J
CI
.E

~ 0
.5

- - -- ti-4-
.......
---- - --
a:
Time in seconds
(c)

Fig. 43. Metronome function: (a) music score; (b) metronome marking
function; (c) graphic score.

The first record describes the function, and P4 gives the initial abscissa
(0), P5 the ordinate at that abscissa (60), P6 the next abscissa (4), P7
the next ordinate (60), etc. The abscissa is in beats and the ordinate in
metronome marking-beats per minute. Successive points on the
function are connected by straight lines, as shown in Fig. 43. As many
segments as desired may be used by putting more points into the SV2
function. The abscissa points need not be uniformly arranged.
The second record tells the Pass II program that a metronome function
is being used and that it starts in variable 50, that is to say, in G(50).
The graphic score in Fig. 43 shows the notes resulting from the
metronome function being applied to the score. The pitch of each note,
plotted against the time it occurs, is shown by the horizontal bars.
Pitch is given on a logarithmic scale, 0 = middle C, + 1 = C 512 • Time
 A SEQUENCE OF TUTORIAL EXAMPLES 89

is in seconds. Such a graphic score has proved to be an effective way of

displaying many computer note sequences.
The first four notes, at a tempo of 60, occupy the first 4 sec. The
second measure is played at an increasing tempo from 4 to 6.5 sec. The
third measure at a tempo of 120 lasts from 6.5 to 8.5 sec. The last two
notes, at a tempo of 30, go from 10 to 14 sec.
The computation relating metronome function to a note's starting
time and duration consists in sampling the metronome function at the
beginning of the note. These sampling points are indicated by ticks on
the function. Thus the sixth note has a tempo value of 75. The duration
and starting time of the note are defined as
Duration = P4 x li sec
= .8 x ~% = .64 sec
and
Starting time = starting time of previous note
+ (P2 of note
- P2 of previous note)· ~%
= 4 + (5 - 4) .~% = 4.8 sec
Because the metronome function for the 13th note is sampled at 12
beats, its value is 30, and a rather long silence occurs between the 12th
and 13th notes. Such a silence is inherent in the algorithm. It is seldom
objectionable; also such large changes in tempo seldom occur.
If several voices are playing simultaneously, the same metronome
function is applied to all.
Metronome functions have proved to be powerful tools for inserting
acce1erandos and ritardandos. Without them, the calculation of starting
times and durations of gradually changing note sequences can be very
tedious. In addition, they enable the composer to write in terms of
conveniently defined beats, rather than seconds, which often turn out
to be unwieldy decimals.

A Note-Generating Subroutine and Graphic Score

The metronome function is built into Pass II. Let us now write a
subroutine PLF3, which will generate a voice from a complete, if
elementary, graphic score. A sample score is shown in Fig. 44. Four
functions are used to describe the voice-duration, duty factor, pitch,
and amplitude. All are functions of time, which in this example goes
°
from to 13 sec.
The duration function in Fig. 44a gives the time from the beginning
of one note to the beginning of the following note. The first note starts
90 CHAPTER TWO
2
<:
o
(a) ;

~ 0 . . . . ..
(hI! :;
o

.<::
·:f---J/
0

.5
u
(c) ~ 0
0..
-.5

-I

60

~ 50

(d) ~0.
E
ct 40

o
Time in seconds

Fig. 44. Graphic score and resulting notes: (a) duration; (b) duty factor;
(c) pitch; and (d) amplitude.

at t = 0, where the value of the duration function is .5 sec. The second

note starts at.5 sec, where the value of the duration function is.8 sec; the
third note starts at 1.3 sec (1.3 = .5 + .8), and so on. In other words,
the duration function is sampled at the beginning of each note to obtain
the interval to the beginning of the next note. The actual sampling times
are shown as dots along the abscissa of Fig. 44a. Although there are
more advanced ways of representing durations graphically, this
representation is easy to program and is useful for a number of
purposes.
The rhythm is represented by the duration function. The' style of
playing-legato or staccato-is represented by the duty-factor function,
Fig. 44b, which gives the proportion of the interval between the starting
times of successive notes that is occupied by sound. The first note has a
duty factor of .5. Its length will be .25 sec (.5 duty factor x .5-sec
interval to start of the second note). A duty factor of .25 produces an
 A SEQUENCE OF TUTORIAL EXAMPLES 91

exceedingly staccato note sequence. A duty factor of 1 produces a

smooth legato. A duty factor greater than 1 causes overlap of successive
notes. Like the duration function, the duty-factor function is sampled.
The duty factor for each note is the value of the duty-factor function at
the starting time of the note. Duty-factor function will be somewhat
arbitrarily used for one other purpose. A negative duty factor will
indicate a rest-the note will be omitted entirely.
The pitch and amplitude functions, Fig. 44c and d, are also sampled
at the starting time of each note to obtain pitch and amplitude. Pitch is
written in our logarithm scale, C262 = 0, CS12 = 1, etc. Amplitude
is written in decibels. A graphic representation of the actual notes that
are generated are shown as horizontal bars superimposed on the pitch
and amplitude functions. Pitches can be read from the pitch scale,
amplitudes from the amplitude scale, durations and starting times from
the beginning and ending times of either the pitch or amplitude bars.
Because of the sampling process, the left end of each bar starts on the
pitch or amplitude function.
The functions shown graphically in Fig. 44 must be represented
numerically in the computer memory in Pass I. The convention that
represented the metronome function will again be employed; successive
breakpoints will be given by their abscissa and ordinate values. The
data will be stored in the Pass I D(2000) array by SVI cards. Thus the
duration, duty factor, pitch, and amplitude functions in Fig. 44 are
described by the following four records

Duration: SVI 0 50 0 .5 3 2 4 .3 8 1 12 .5 13 .5
Duty factor: SVI 0 65 0 .5 3 .5 4 1 6.7 1 7 - .25 8 - .25
8.3 1 13 1 ;
Pitch: SVI 0 85 0 0 3 1 7.5 -1 13 0
Amplitude: SVI 0 95 0 40 7.5 60 13 40 ;

The functions start in D(50). The amount of memory occupied by a

function depends on the number of breakpoints. Successive functions
have been arbitrarily spaced by sufficient multiples of five so as not to
overlap.
A F0RTRAN function C0N has been provided to read the graphic
functions. The statement
Z = C0N(D, N, T)
sets Z equal to the value, at time T, of the function that starts at D(N).
92 CHAPTER TWO

Thus, for example,

Z = C0N (D, 95, 3.0)
would set Z equal to 48, which is the value of the amplitude function at
3 sec. The values are computed by interpolating a straight line between
the breakpoints that surround T (0 and 7.5 in the specific example).
C0N must search the D array to find these breakpoints. It is essential
not to ask for values of the function outside the range of breakpoints
that have been defined. Otherwise C0N may never terminate its
search.
We can now write a subroutine PLF3 to generate notes from graphic
scores. The data record to call this routine is
PLF 0 3 TS END NA NP NDR NDF IN ;
where TS is a time shift giving the starting time of the sequence of notes
to be produced by PLF3; END is the duration of the sequence;
NA, NP, NDR, and NDF give the starting points in the D array of the
amplitude, pitch, duration, and duty-factor functions, respectively.
IN gives the instrument number.
An example to produce the notes shown in Fig. 44 is:
PLF 0 3 0 13 95 85 50 65 4 ;
We shall assume that the program writes out records in the f6rm (which
has been used frequently)
N0T TS IN D AMP PITCH ;
where AMP and PITCH are in decibels and log units.
The annotated PLF3 subroutine follows.
Text Notes

SUBR0UTINE PLF3
C0MM0N IP, P, D
DIMENSI0N IP(10), P(100), D(2000)
TS = P(4)
END = P(5)
NA = P(6)
NP = P(7)
NDR = P(8)
NDF = P(9)
P(1) = 1.0 2
P(3) = P(10)
IP(1) = 6
T = 0.0 3
 A SEQUENCE OF TUTORIAL EXAMPLES 93

100 DR = C0N(D, NDR, T) 4

IF(T + DR - END) 101, 101, 104 5
101 P(2) = T + TS 6
P(4) = DR * C0N(D, NDF, T) 7
IF(P(4)) 103, 103, 102 8
102 P(5) = C0N(D, NA, T) 9
P(6) = C0N(D, NP, T)
CALL WRITE1(10)
103 T = T + DR 10
G0 T0 100 11
104 RETURN
END

Notes
1. These statements extract the essential information for the PLF3
from the P array.
2. These statements set the constant parts of the P array in preparation
for writing out N0T records, and they set the word count.
3. T is the starting time of the next note to be generated (not including
the time shift TS). It is set initially at zero and computed as a
running variable and is increased by the interval between successive
notes after each note is generated. T is also the variable used to
specify abscissa values in C0N.
4. DR is the interval between successive notes as obtained by C0N
from the duration function.
5. This statement checks to see whether the starting time of the next
note is greater than the ending time, END. If so, the current note is
not generated and PLF3 is terminated.
6. The time shift TS is added to T to obtain the starting time of the
note.
7. The duration of the note is computed as the interval times the duty
factor.
8. This statement checks for a rest. If the duration is zero or negative,
owing to a zero or negative duty factor, no N0T is written out and
the program proceeds to the next note.
9. These statements compute the rest of the N0T parameters and
write out the N0T record.
10. This statement adds T to the starting time of the next note.
11. This statement transfers control in order to generate the next note.

Several features of the operation of PLF3 may be pointed out in

Fig. 44. The first four notes are staccato, having large silences between
94 CHAPTER TWO

notes. The next six notes are legato, with no silent intervals between
notes. Two possible notes have been omitted to form a rest.
In terms of the number of notes generated, PLF3 is very efficient.
One PLF3 call produced 16 notes. It could just as well have produced
1600. In contrast to conventional scores, the notation for duration has
the advantage that a second's worth of fast notes requires no more
effort to describe than a second's worth of slow notes. Also ritardandos
and accelerandos are easy to describe by lines with increasing or
decreasing slopes, as illustrated from 4 to 8 and from 8 to l3 sec. Such
tempo changes can have striking acoustical effect.

Pass n Subroutines-PLS Pitch-Quantizing Example

To complete the discussion of N0T-generating subroutines and
note-manipulating subroutines, we will write one second-pass sub-
routine. Pass II routines cannot be used to generate additional notes
since all the N0T records have been carefully sorted in increasing order
of action times, and the addition of more N0T records would disrupt
the ordering. However, PLS routines can change the values of note
parameters (except action times). Since notes of all voices are sorted
together, it is convenient for PLS to embody relations involving several
voices at a particular time. For example, PLS could well be used to
adjust the pitch intervals between voices.
We will not attempt quite as complicated an example as interval
control. Instead we will control the pitches of a single voice so that they
fall exactly on the steps of a previously specified scale. Such a process
makes sense when applied to the output of the PLF3 routine that was
presented in the preceding section. The pitches so generated are samples
of a continuous pitch function and can fall anywhere. Sometimes it is
desirable to limit the possible pitches to a prespecified set or scale. The
scale need not correspond to any known or standard musical scale,
such as ajust scale or a 12-tone scale. An octave can be divided into any
number of intervals; the intervals can be even tempered (equal) or
unequal in size.
Figure 45 gives an example of the output of the routine to be written.
It is applied to the notes generated by the PLF3 program. The pitch
function and notes from Fig. 44 have been redrawn in Fig. 45. For the
scale the octave is divided into five equal intervals, as shown in Fig. 45.
Since pitch is in logarithmic units, these units correspond to equal
musical intervals. The PLS routine will adjust the pitches of the notes
generated by PLS to the closest scale step. The pitches generated by
 A SEQUENCE OF TUTORIAL EXAMPLES 95

1.0
Scale tones

.5

-.5

-Generated pitch
--- Quantized pitch
-1.0

I I I I I I I I I I I I I
o 2 3 4 5 6 7 8 9 10 II 12 13

Time in seconds

Fig. 45. Pitch quantizing by a PLS routine.

PLF are shown as solid horizontal bars, and the adjusted pitches are
shown as dashed bars. The adjustment may be either up or down,
depending on which scale step is closest. The process of adjustment is
called pitch quantizing.
Note that the first and sixth notes happened to fall exactly on a
scale step and require no quantizing. Also the last two pairs of notes
become pairs of repeated notes as a result of quantizing. Quantizing
tends to produce repeated notes if the scale steps are large and the change
in pitch between successive notes is small.
In order to write a PLS routine, it is necessary to understand a few
details of the operation of Pass II. All the data records in a section are
read into a large array D(10,000), 10,000 locations long in the training
orchestra. An array 1(1000) is computed by sorting so that
1(1) = the address in D of the beginning of "first" data record
where "first" means smallest action time
1(2) = the address in D of the beginning of "second" data
record
etc.
96 CHAPTER TWO

For example if the thirteenth data record is

N0T 19 2 4 60 .167 ;

and is stored starting at D(109), then

1(13) = 109

and
D(109) = 6 (Word count)
D(IlO) = 1 (N0T == 1)
D(IlI) = 19
D(112) =2
D(1l3) = 4
D(II4) = 60
D(1l5) = .167
After 1(1000) is computed, the program goes through the data
records in order of increasing action times, executing any PLS routines,
storing any SV2 data in the G(1000) Pass II data array, and writing out
N0T records with the aid of the C0NVT subroutine.
A PLS function can modify any N0T records with action times
greater than the action time of the PLS function. It cannot affect N0T
records with action times less than the PLS function, since these will
already have been written before PLS is executed. The Pass II memory
G(1000) will contain the numbers from any SV2 cards with action
times less than the action time on the PLS function; it will not contain
any data from SV2 cards whose action times are greater than on the
PLS function.
The scale will be stored in Pass II memory by a SV2 statement giving
the number of steps in the scale, followed by the pitches of these steps.
Thus the scale used in Fig. 45 is inserted in the memory by the record
SV20 100 11 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 ;
These data will go into G memory at time O.
The PLS 1 function will be called at action time 0 by the statement
PLS 0 1 100;
where P3 = 1 indicates PLS 1 and the 100 gives the starting point of the
scale in the G array.
An annotated program to carry out the pitch quantizing follows.
 A SEQUENCE OF TUTORIAL EXAMPLES 97

Text Notes

SUBR0UTINE PLS1
C0MM0N IP, P, G, I, T, D 1
DIMENSI0N IP(10), P(100), G(1000), 1(1000),
T(1000), D(10,OOO)
Il = IP(2) 2
IN = IP(3) 3
NQ = D(1l + 4) 4
NB = NQ + 1
NL = NQ + IFIX(G(NQ))
D0 103 J = 1, IN 5
ID = I(J)
IF(D(lD + 1) - 1.0) 103, 100, 103 6
100 FREQ = D(ID + 6) 7
MIN = 1,000,000.0 8
D0 102 K = NB, NL
IF(ABS(FREQ - G(K)) - MIN) 101, 102, 102
101 MIN = ABS(FREQ - G(K)
QFREQ = G(K)
102 C0NTINUE
D(lD + 6) = QFREQ 9
103 C0NTINUE
RETURN
END

Notes
1. This common statement and the subsequent dimension statement
describe the main data arrays in Pass II and must agree with the
corresponding statements in the Pass II main program. IP gives
certain miscellaneous constants, P is the communication array
from which data records are read and written, G is general variable
storage, I indexes the D array in action-time order, T contains
action times and is primarily used in the sorting process, D contains
the data records.
2. When PLS is called, IP(2) contains the address in the D array at
which the PLS data record is located. In this case if
IP(2) = 27
then
D(27) = 4 (Word count)
D(2S) = 10 (PLS == 10)
D(29) = 0
D(30) = 1
D(3l) = 100
 98 CHAPTER TWO

3. IP(3) contains the number of data records in D. The main D0

loop in the PLS routine will examine I(J) for J = 1 to IP(3).
4. This and the subsequent two statements determine NB and NL as
the first and last locations of the scale description in the G array. A
D0 loop will test these locations.
5. The main D0100p examines all data statements in order of ascending
action times. ID in the sub seq uent statement is the D address of the
data statement.
6. This statement checks to see if the data statement is a N0T record.
If not, it is skipped; if so, the pitch variable P6 is quantized.
7. FREQ is set equal to the pitch P6.
8. This and the subsequent statements to 102 determine the scale step
that is closest to FREQ. MIN is initially set to a very large value.
The absolute value of (FREQ - each scale step) is compared with
MIN and if it is smaller than MIN, MIN is reset to that value. In this
way MIN ends being the smallest interval and QFREQ ends being
the closest frequency.
9. This statement resets the pitch D(ID + 6) to the closest scale step.

The PLS routines tend to be both longer and logically more compli-
cated than the PLF routines. The steps in the example just discussed are
typical. Actually, they were not all necessary for the problem at hand.
The pitches could have been quantized by the PLF routine as they were
generated. Even if the quantizing were done in Pass II, it would not
have been necessary to go through the D array in order of action times.
However, for slightly more complicated operations, such as quantizing
the intervals between voices, all the Pass II steps are essential.
Another simplification in the program consists in writing out the
scale for all the octaves in which it is to be used. In many cases, only
one octave is written out; the actual pitches are translated to this
octave before being quantized; and the quantized pitches are translated
back to their original octave. The possibilities open to the composer
are almost endless.

Interactions Between Instruments

The final process to be considered in this chapter involves interactions
between instruments. The desirability for such interactions arises from
\
the limitations of the "note concept," which defines sounds as having
~ starting and ending times. Sometimes it is desirable to produce con-
tinuous sounds that change from time to time in controlled ways. As
 A SEQUENCE OF TUTORIAL EXAMPLES 99

we shall show, this can be done by using the output of an instrument as

an input to another instrument. The first instrument is degenerate in the
sense that it produces no acoustic output. Instead it plays a series of
notes that generate a long and frequently changing modulation function
for the second instrument. The second instrument may play only a
single long note whose sound is varied by the parameter supplied by
the first instrument.
A typical and important use of interactions is amplitUde control to
produce swells and diminuendos as notated on the conventional score
in Fig. 46. Such a modulation is unwieldy to program with the apparatus
previously described. Although we can draw continuous amplitude

I J J J Jl
=
(0 )

V3 V4 V5 V6 P5 82 P9 PIO PII 83 P7 VI V2

'~2LSG

82
r6

Instrument
#1
¥ 83

Instrument
#2
(b)

-£
~ .5~-------'

Instrument
Time in seconds #3

(c)

Fig. 46. Interconnected instruments for amplitude and glissando control:

(a) musical score; (b) block diagrams of instruments; (c) continuous control
functions.
100 CHAPTER TWO

functions, as is done in Fig. 44, these are sampled at the beginning of

each note and that amplitude is held constant for the duration of the
note; this effect is clearly not the desired objective of Fig. 46. Further-
more, since the inputs to instruments are set at the beginning of each
note, the sampling effect is hard to avoid. In Fig. 46a, measures 4 and
5 also call for a combination of glissando with amplitude variation. The
last measure applies a continuously changing amplitude control to a
sequence of notes. As an example of interacting instruments, we shall
produce these effects.
The apparatus for interactions is the input-output blocks BI-BIO,
which are shared by all instruments. The output of instrument A may be
left in a block for subsequent use by instrument B. This requires that A
be computed before B. Pass III computes the instruments in order of
their numbers, 3 any that are #1 first, then any that are #2, and so forth.
Hence by making A a lower numbered instrument than B, the proper
order can be guaranteed. An additional requirement is that the block
used for communication cannot be used for other purposes which
would overwrite the output of A before B uses it. Also, in contrast to
most instruments, A can generate only one voice at a time.
A special unit generator LSG, which rapidly computes functions
formed from straight-line segments, is useful in instruments that
generate control functions. Two such generators are used for instru-
ments I and 2 in Fig. 46 to produce amplitude- and frequency-control
signals, respectively. The operation of LSG is simply

111 = 111 + 121

0 1 = III
or in other words 11 is incremented by 12 for each sample and 0 is
equal to 11. Because only addition is involved, the process is rapid. 11
will be set to a desired initial value and 12 to the slope of a linear
function that starts at I I. I I and 12 can be reset at any time, thus
changing the value of 0 and the slope abruptly. In instrument #1, Pass
III variables 3 and 4 are used for 11 and 12. These will be set with SV3
records which are generated by a Pass I subroutine PLF4 to achieve a
particular amplitude-control function. Instrument #2 produces the
same effect for pitch. The outputs of instruments 1 and 2 are put in
blocks B2 and B3 where they form inputs to instrument 3.
Instrument 3 is a modification of the envelope instrument which was

3 As of February 21, 1968, this feature was not yet programmed in Music V.
However, it seems both desirable and easy to insert.
 A SEQUENCE OF TUTORIAL EXAMPLES 101

developed in Fig. 39, and it uses the C0NVT function for that instru-
ment. The additional amplitude function B2 is multiplied by the normal
amplitude input P5. The continuous amplitude-control function is
written in decibels (as shown in Fig. 46c) , and B2 is the exponential
transformation
B2 = 10 ** (Continuous ami~itude functiOn)

Thus, the decibels of the normal amplitude function and the continuous
function are additive. If the continuous function is 10 dB and the normal
function is 50 dB, the resulting sound will be at 60 dB.
The normal frequency input P7 is mUltiplied by the additional
frequency-control function B3. The continuous pitch function (also
shown in Fig. 46c) will be written in our standard logarithmic scale, and
B3 will be the exponential transformation
B3 = 2 ** continuous pitch function
Thus a continuous pitch function of 0 produces no change in pitch, a
continuous pitch function of 1 produces a one-octave upward shift, and
so forth. The computation of V3-V6 to achieve both the exponential
conversions and the proper increments will be done by a PLF4
subroutine.
Input VI specifies the proportion of frequency shift in the vibrato,
proportionality being controlled by a multiplier. Such Pass III multi-
plication is essential rather than multiplication by the C0NVT function,
because frequency can vary over a note.
The annotated PLF4 program is given below. The pitch and ampli-
tude functions will be stored as Pass I variables in the usual notation.
The functions shown in Fig. 46c are stored by the statements
SV1 0 50 0 0 4 20 8 0 12 20 19.99 0 20 20 24 0 ;
SVI 0 70 0 .583 12 .583 14 1.167 16 .333 19.99 .583 20 0
240;
The calling record for PLF4 is
PLF 0 4 TS END FA FP ;
where TS is the starting time of the control functions, END is the
duration of the control functions, FA is the starting variable of the
amplitude function, and FP is the starting variable of the pitch function.
For the example
PLF 0 4 0 24 50 70 ;
is the specific calling record.
102 CHAPTER TWO

PLF4 generates a sequence of SV3 records to form the inputs to

instruments 1 and 2, and generates two N0T records to activate these
instruments from 0 to 24 sec.

Text Notes

SUBR0UTINE PLF4
C0MM0N IP, P, D
DIMENSION IP(10), P(100), D(2000)
TS = P(4)
END = P(5)
NA = P(6)
NP = P(7)
1= NA 2
IP(1) = 5
P(l) = 4.0
P(3) = 3.0
100 P(4) = 10.0 ** (D(I + 1)/20.0) 3
P(5) = (10.0 ** (D(I + 3)/20.0) - P(4»/«D(I + 2)
- D(I» * D(4»
P(2) = TS + D(I) 4
CALL WRITE 1(1 0)
IF (D(I + 2) - END) 101, 102, 102 5
101 1=1+2 6
G0 T0 100
102 1= NP
P(3) = 5.0
103 P(4) = 2.0 ** D(I + 1)
P(5) = «2.0 ** D(I + 3) - P(4»/«D(1 + 2)
- D(I» * D(4»
P(2) = TS + D(I)
CALL WRITEl(10)
IF(D(1 + 2) - END) 104, 105, 105
104 1=1+2
G0 T0 103
105 IP(1) = 4 8
P(1) = 1.0
P(2) = TS
P(3) = 1.0
P(4) = END
CALL WRITE1(10)
P(3) = 2.0
CALL WRITE1(10)
RETURN
END
 A SEQUENCE OF TUTORIAL EXAMPLES 103

Notes
1. These statements extract the essential information for PLF4 from
the P array.
2. These statements prepare to write SV3 records for V3 and V4.
P(I) is 4 for SV3. P(3) = 3.0 designates V3 as the first variable.
One pair of V3 and V4 values will be written for each segment of the
amplitude function. I = NA will set the initial value of the equations
starting at 100 for the first segment.
3. This and the subsequent line calculate the initial value and slope for
the first segment. The slope is in units per sample. D(4) is the
sampling rate.
4. The time of the SV3 card is the beginning time of the first segment
plus TS.
5. This statement terminates the amplitude function at the end of the
current segment if D(I + 2) ~ END.
6. I is incremented by 2 and control is transferred to 100 to continue
with the next segment.
7. These statements write out SV3 records for variables 5 and 6 to
produce the pitch control. The process is exactly analogous to
amplitude control.
8. The rest of the program writes out two N0T records

N0T TS 1 END
N0T TS 2 END

that play two notes on instruments 1 and 2 which start at TS and

have duration END.

The score records to produce the Fig. 46 output are given below. The
definition of the instruments and the Pass III stored functions are
omitted since they are completely standard.

SVI 0 50 0 0 4 20 8 0 12 20 19.99 0 20 20 24 0 ;
SVI 0 70 0 .583 12 .583 14 1.167 16 .333 19.99 .583 20 0
240;
PLF 0 4 0 24 50 70 ;
N0T 0 3 11.8 40 262 ;
N0T 12 3 7.8 40 262 ;
N0T 20 3 .8 40 392 ;
N0T 21 3 .8 40 349 ;
N0T 22 3 .8 40 330 ;
N0T 23 3 .8 40 294 ;
104 CHAPTER TWO

.g 60

i. 50
E
<X
40

1.5

.£:
o
a::

H
o

o 2 4 6 8 10 12 14 16 18 20 22 24
Time in seconds

Fig. 47. Graphic score with continuous changes in amplitude and pitch.

The graphic score of the resulting sound is shown in Fig. 47. Beginnings
and ends of individual notes are indicated by short vertical bars on the
amplitude and pitch curves. An attack or decay will be produced by the
envelope generator at these times. Amplitude and pitch changes occur
continuously and independently of note boundaries.

Parting Exhortations to the Student

The tutorial examples are now complete. However, the student's
task-harnessing the computer to his objectives-has just begun. A
mere reading of the examples is not sufficient to master their content.
The examples are a far from complete description of the Music V
program; the Music V program as written will produce only a fraction
of the wanted and achievable computer sounds.
Programming skills come only with practice. The problems accom-
panying this chapter provide some possibilities for practice. Their
solution by the student is greatly recommended. If limited time forces a
choice between reading the chapter and working the problems, working
the problems is to be preferred.
Most of the problems are based on the material given in the chapter.
In some cases, more details about the Music V program must be
obtained. These can be found in the Music V Handbook which
 A SEQUENCE OF TUTORIAL EXAMPLES 105

forms the next part of this book. The handbook is intended to be a

complete description of the program, arranged and indexed for
references. Information about Music V that has been given in this
chapter is also presented in the handbook, and is usually easier to find
there. The student should become accustomed to answering his
questions from the handbook as soon as possible.
Although it is not necessary to read the entire handbook to make use
of its information, anyone who plans to make extensive use of Music V
should read the sections describing the operations of the various parts
of the programs. Detailed block diagrams as well as verbal descrip-
tions of operations are included. Reading the handbook is a helpful
preparation for reading the programs themselves.
Music V is written almost entirely in F0RTRAN. Consequently, it is
practical to read the programs and understand their operation. Such
understanding is essential if major modifications of the programs are to
be made. The advanced user will want to make such modifications;
Music V was written with this objective in mind. Hence the student's
final teacher, and the final arbiter of questions about the operation of
Music V, is the programs. Such is the nature of computer programs.

Annotated References by Subject

Computers in General
J. Bernstein, The Analytical Engine (Random House, New York, 1964). A non-
mathematical and elementary introduction to computers and what they
can do.
A. Hassitt, Computer Programming and Computer Systems (Academic Press,
New York, 1967). A discussion of programming from an elementary to an
advanced viewpoint.
Fortran Programming
S. C. Plumb, Introduction to Fortran (McGraw-Hili, New York, 1964).
E. I. Organick, Fortran IV Primer (Addison-Wesley, Reading, Mass., 1966).
S. V. Pollack, A Guide to Fortran IV (Columbia University Press, New York,
1965).
These are three self-instructional texts that teach Fortran.
Graphic Scores
M. V. Mathews and L. Rosler, "Graphic Scores," Perspectives of New Music 6,
No.2 (1968). A detailed article illustrating one technique for composing
with the aid of a computer.

Problems for Chapter 2

Parameters of Training Orchestra
Sampling rate-20,OOO Hz
Function block length-512
Number of functions-IO
106 CHAPTER TWO

Range of functions- - 1 < F < + 1

1-0 block length-512
Number of 1-0 blocks-l0
Range of unit generator inputs and outputs- - 2047 to + 2047
Maximum number of note parameters-30
Number of Pass III variables-200
Maximum number of voices-30
Pass II G array length-l 000
Pass I D array length-2000

Even-Tempered Scale
Frequency Logarithmic
Note in hertz pitch

C 262 0
C# 277 .083
D 294 .167
D# 311 .250
E 330 .333
F 349 .417
F# 370 .500
G 392 .583
G# 415 .667
A 440 .750
A# 466 .833
B 494 .917

Introductory Score- Writing Problem

1. Using the orchestra defined in Fig. 27 write the computer score for the
following conventional score.

120 ri t (J 100)

p mf f

Assume that in amplitude, p ~ 50, mf ~ 150, and f ~ 500. Assume that

staccato notes sound for .5 the nominal time occupied by the note (for
example, a staccato quarter note at a tempo of 120 would sound for .25
sec). Legato notes sound for .8 of their nominal time, and slurred notes for
1.1 of their nominal time. (Remember that a Music V instrument can play
 A SEQUENCE OF TUTORIAL EXAMPLES 107

more than one note at a time. The limit in the training orchestra is 30
simultaneous voices.)

Simple Unit Generators

2. Write out the samples Fn(j) j = o... 511 for the following stored
functions. To shorten your answers, use ... to indicate a sequence of
identical samples.

(a) F4, a symmetrical square wave with amplitude + 1 or - 1

+1

-I

(b) F5, a triple pulse wave as shown

o
(c) F6, a sine wave of peak amplitude 1 (write only the first 20 samples)

(d) F7, an attack function with shape

1
o 5
\
506 511

3. For an oscillator with

11 = 500
12 = 50.35
and a function FI
108 CHAPTER TWO

o 500 511

calculate S1, [Sl]MOd 511, F([SI]MOd 511)

and 0 1 for i = 0 ... 15
Assume [SI]MOd 511 is rounded to the next lower integer in looking up values
of Fl. Calculate the truncation error in 0 1 due to rounding for i = o... 15.
4. Instrument 1 consists of only one oscillator
0SC P5 P6 B2 F4 P30 ;
where F4 is the symmetrical square wave defined in problem 2a. Write the
samples B2(l) ... B2(20) generated by the following notes
(a) N0T 0 1 .001 1000 250 ;
(b) N0T .002 1 .001 1000 50 ;
(c) N0T .004 1 .001 500 128.3 ;
(d) N0T .006 1 .001 1000 600 ;
(e) The numerical frequency of the last note is
600
511 ·20,000 = 23,500 Hz

This frequency is much greater than half the sampling rate. What is the
apparent period of B2(1) ... B2(20)? This period (about 6 samples) results
from foldover.
5. Instrument 1 shown in Fig. 27 uses the symmetrical square wave of
problem 2a for its stored function. Write the output samples So, Sb ... ,
S60,OOO resulting from the following score. Abbreviate your answer by
designating blocks of zero samples by ...
N0T 0 1 .0005 1000 60 ;
N0T .5 1 .0006 500 200 ;
N0T 1 1 .0002 100 10 ;
N0T 2 1 .001 500 70 ;
N0T 2.0002 1 .0004 500 100
6. Instrument 2 shown below uses F4 function of problem 2a. It plays
the note
N0T 0 2 .001 500 80 3 100 ;
Plot the samples
B2(1) ... B2(20)
B3(1) ... B3(20)
B4(l) ... B4(20)
 A SEQUENCE OF TUTORIAL EXAMPLES 109

P5 P6 P7 P8

Instrument 2

Simple Instruments
7. Score the instrument diagrammed here.

(a) What do P5, P6, and P7 control?

(b) What is the % vibrato?
(c) <Jo amplitude variation?
(d) What is the rate of vibrato?
110 CHAPTER TWO

(e) Write score records for Fl, an attack and decay function; F2, a
vibrato function; and F3, a modified square-wave waveform.
(f) Write the score for the following passage

J'J .r~----
J I J j J I J.
8. Diagram, score, and write functions and a note for. an instrument
that has attack and decay in amplitude and a frequency attack. The
frequency of each note should start 10% low and rise linearly to the final
frequency of the note within the first 1070 of the note's duration.
9. Diagram and score an instrument with attack and decay in amplitude,
with vibrato, and with attack and decay on the vibrato.
10. Diagram an instrument that uses four 0SC's to change the wave-
form of a note as a function of both amplitude and frequency. The com-
position of the output waveform should be
A·[{1000 - f}{(1 - A)0SC 1 + A·0SC 2 }
+ {f}{(1 - A)0SC 3 + A·0SC4 }]
where A is an amplitude control going from 0 to 1 and f is frequency in
hertz.

CfJNVT Functions
11. Write a C0NVT function for the instrument shown in Fig. 32 which
will process a note record of the form
N0T T 2 D A F ;
where A is amplitude in decibels and F is frequency in hertz. V50 is the
proportion of vibrato. For each N0T record, C0NVT should write out
three records to produce a three-note chord, the highest voice having a
frequency of A Hz, the middle voice A/2 Hz, and the lowest voice A/4 Hz.
12. Write a C0NVT function for the instrument shown in Fig. 33 which
reads a N0T record of the form
N0T T 3 D Al A 2 •• • An Freq
where A1 ... An is a sequence of amplitudes in decibels and Freq is frequency
in hertz. The C0NVT function outputs n + 1 successive notes of equal
duration, whose total duration is D. The first note starts at amplitude 0
(linear scale) and ends at A1(dB), the second goes from Al to A2 , ••• , the
last goes from An(dB) to 0 (linear scale).

Additional Unit Generators

13. Design an instrument with an amplitude-modulated band-pass noise
having a bandwidth equal to t the center frequency of the noise band, and
a noise band whose center frequency changes linearly from an initial to a
final frequency during each note.
 A SEQUENCE OF TUTORIAL EXAMPLES 111

14. Design an instrument having a random amplitude variation of

± 50'70 of the average amplitude and a low-pass spectrum going from 0 to
15 Hz.
15. Design an instrument producing a band-pass noise by both frequency
and amplitude modulation. Have the center frequency of the noise band
controlled by P7 and the (bandwidth/center frequency) ratio by P8.
16. Write a C0NVT function for the instrument shown in Fig. 39
which will generate notes with an attack time of .1 sec and a decay time of
.2 sec, provided the note duration is greater than .3 sec. The steady-state
time should be (duration - .3) sec. For notes of duration between .2 and
.3 sec, the attack time should be .1 sec and the decay time (duration - .1)
sec. Any durations less than .2 sec should be increased to .2 sec.
17. Compute 12 and 13 for filters with a center frequency of 500 Hz and
bandwidths of 2 Hz, 10 Hz, 50 Hz, and 260 Hz. What is the dc gain of these
filters? What is the peak frequency gain? What is the maximum input
signal that will not cause the output to exceed 2048 ?

Composing Subroutines
18. Write a set of PLF routines that will process note data in Pass 1
memory. Assume that the note data are stored in the Pass I D array in the
manner used for the Fig. 41 example, and that notes will be written for the
instrument shown in Fig. 39. Write the following subroutines:
(a) PLFI rewrites n notes in the D array, multiplying all logarithmic pitch
intervals by S, adding a constant K to the logarithmic pitch intervals, and
changing the tempo by a factor T.
(b) PLF2 substitutes a new note for note n in the array.
(c) PLF3 makes a copy of n notes starting at D(m) and stores the copied
notes at D(p), overwriting anything that was previously at D(p).
(d) PLF4 divides each of n notes starting at D(m) into k notes of equal
length whose total duration equals that of the note they replace. The new
notes are written starting at D(p).
(e) PLF5 writes N0T records for n notes starting at D(m). The starting
times of all notes are shifted by T sec.
Use these subroutines to compute a composition.

Graphic Scores
19. Write a subroutine PLFI that will generate pitch and amplitude
functions as the computed functions
Pitch(t) = f 1 (t) * f 2 (t) + f 3 (t) * f 4 (t)
Amplitude(t) = f5 (t) * f6 (t) + f7(t) * fs(t)
where f 1 (t) through fs(t) are functions stored in the D array. Compute the
starting and stopping times of notes as the positive-going zero crossings and
the negative-going zero crossings, respectively, of a function
Notes(t) = fg(t) * f10(t) + (1 - fg(t)) * f12 (t)
112 CHAPTER TWO

where fg(t) , flO(t) , f12(t) are stored in the D array. Let f 1o(t) and f12(t)
correspond to the rhythmic sequence of two well-known melodies. What
notes will be generated when fg(t) = 1 ?; when fg(t) = O?; when fg(t) has
some intermediate value? Follow the general procedures used in the Fig.
44 example.

Pitch Quantizing
20. Write PLS1, a pitch-quantizing routine which will quantize a voice
for instrument 1 into the closest note in the C major scale. Assume that
voices for instruments 2 and 3 produce notes in synchrony with instrument
1. Adjust these voices to harmonize instrument 1 according to the following
rules.
(a) Harmonize C and E with the chord CEG.
(b) Harmonize F and A with the chord FAC.
(c) Harmonize Band D with GBD.
(d) Harmonize G with CEG if it starts on a multiple of four beats and
with GED if it starts on any other beat.
Use a minimum adjustment of the other voices to achieve these chords.

Interconnected Instruments
21. Define an orchestra and an appropriate C0NVT function so that
the output of an instrument is the sum of two 0SC's, the proportion of
each being determined by two separate instruments 11 and I2. The propor-
tion will change continuously and frequently during the course of the notes
to add interest to the sound quality. Use LSG unit generators and follow
the general procedures of the Fig. 46 example.
3 Music V Manual
M. V. Mathews, Joan E. Miller,
F. R. Moore, and J. C. Risset
3 Music V Manual

1. Introduction
This chapter contains a detailed description of the operation and
structure of the Music V program. It provides reference material for
users of Music V and source material for those who desire intimate
knowledge of a sound-generating program in order to write their
own.
Music V is the direct descendant of Music IV, a program that was
widely used for five years and has been described in the literature. 1
Music V had to be rewritten to change from a second to a third genera-
tion computer (the IBM 7094 to the GE 645). However, in the process
certain improvements were made, especially changes that made the
program more easily adapted to other computers. It may be helpful to
'list these changes for the benefit of users of Music IV.

Principal Differences between Music IV and Music V

1. Music V is written almost entirely in F0RTRAN IV; it is much
easier to use on a wide variety of computers. In addition, the F0RTRAN
programs have been written so as to be easily modifiable to accommo-
date the different memory sizes and different word lengths of various
computers.
1 See Annotated References at end of chapter.

115
116 CHAPTER THREE

Despite being written in F0RTRAN, Music V is potentially as fast

as Music IV. This potentiality can be achieved by writing the inner
loops for certain computations (the unit generators) in basic machine
language. Such programs are, of course, specific to a given computer,
but at most only a few hundred instructions are involved.
F0RTRAN unit generators can be intermixed with basic machine-
language generators. Initially, the program can be operated entirely
with F0RTRAN generators. Gradually, the most frequently used
generators can be coded in machine language. Exotic and infrequently
used generators may remain in F0RTRAN at little cost. New generators
can easily be added in F0RTRAN.
2. Instruments are defined as part of the score rather than in a
separate program. (In Music IV the orchestra was assembled by the
BE F AP assembly program.) In this way the entire composition-notes
and timbres-is specified in a single document, the score. In addition,
instruments may be redefined or changed at any point in the score.
3. A given instrument may play any number of voices simultaneously.
Only one instrument of a given type need be defined; the composer no
longer need worry about losing notes that overlap in time on an
instrument. Unit generators are also mUltiply used; only one copy of
each type of generator is in the memory; memory is thus conserved.
4. A free-field format for score cards is used. Successive fields are
separated by one or more blanks or by commas. Mnemonics are used
to denote operation codes and unit-generator types. This form of score
is easier both to write and to read than the Music IV fixed-field score.
The score is interpreted by a completely separate subroutine READ I
and the output is entirely in numerical form. Therefore, it is possible
to change the form of the score simply by replacing the READ I
routine. Moreover, since all subsequent parts of the program are strictly
numerical, a maximum of machine independence is achieved in the
F0RTRAN programs.

Overview of Music V
A block diagram of the over-all operation of the programs is shown
in Fig. 48. The main programs, the principal subroutines, the flow of
control, and the flow of data are indicated. The few basic machine-
language programs are especially marked.
Pass I causes the score to be read by the READ I subroutine. The
score may be thought of as a sequence of data cards prepared by the
user, although the actual medium could also be a computer-connected
typewriter, a graphic computer, or a data file.
 MUSIC V MANUAL 117

Key:
------~.. ~
Flow of control Flow of Optional F(IlRTRAN IV Machine-language
in program da ta branch subroutine subroutine

Fig. 48. Block diagram of Music V operation.

Cards are processed by Pass I in the order in which they occur in the
score. Data are grouped into data statements which are terminated by a
semicolon; a data statement need not correspond to a single card. The
first field of the data statement specifies an operation code, and the
second field specifies an action time when the operation is to be done.
This time is measured from the beginning of each section of the music.
The other fields may vary depending on the particular operation code.
118 CHAPTER THREE

The total number of fields may vary; no more than necessary need be
used.
The principal operations are to

(1) Cause a note to be played

(2) Define an instrument
(3) Store data in Pass I, II, or III memory
(4) Call a subroutine in Pass I or II
(5) Generate and store a function in Pass III
(6) Terminate a section or a composition.

Pass I calls several subroutines. The function of the READ and

ERR0R subroutines are obvious. The PLF subroutines are note-
processing and generating routines which the composer has the option
to provide if he wishes to make use of this possibility. M0VR and
M0VL are two short machine-language routines that move a character
to the right and left end, respectively, of the computer word. These are
two of the few essential machine-language routines that must be
provided.
Data statements are sent to Pass II via a data file recorded on disk
or tape. Each statement is still labeled with an action time in the second
field. The principal function of Pass II is to sort the data statements
into ascending order of action times. (In Pass I, action times need not
be ordered; in Pass III a strictly ascending order is required.) The
sorting is carried out by two subroutines, S0RTFL and S0RT. These
are provided as F0RTRAN IV routines; however, the sorting process
can be substantially speeded by writing or obtaining machine language
versions. Sorting programs are quite generally available.
After sorting the data statements for time, Pass II (optionally) applies
a metronome function to distort the time scale. Subroutine C0N is
used to read the metronome markings which are stored in the Pass II
memory. Gradual accelerandos and ritardandos are possible, as well as
sudden changes in tempo.
User-provided subroutines, called PLS subroutines, may be
optionally supplied and applied to the data records after time sorting.
Just before each data statement is sent to Pass III, a C0NVT sub-
routine operates on all its fields. C0NVT must be supplied by the user;
it replaces all the CVT routines in Music IV. For example, it is often
given the job of converting frequency notation from some humanly
simple scale like 12 tones-I ... 12-to the proper input numbers for
oscillator frequency control. Inputs for attack and decay generators are
 MUSIC V MANUAL 119

conveniently computed here. Frequently, C0NVT adds parameters to

the data statement.
The actual acoustic samples are computed in Pass III. The unit
generators are encoded in SAMGEN (in basic machine language) and
F0RSAM (in F0RTRAN IV). The Pass III program organizes these
unit generators into instruments and plays the instruments as specified
by the score. In addition, the GEN routines may be called upon to
compute functions that are stored in the Pass III memory and are
referred to by unit generators (e.g., 0SC).
Data statements which are the input of Pass III have action times
written in their second field; these action times are now monotonically
ordered; they determine the times at which all processing and generating
in Pass III are performed.
Almost all information in Pass III is stored in one large array called
1. It contains instrument definitions, parameters of notes currently
being played, stored functions (from GEN routines), input-output
blocks for unit generators, and certain other data. The size of I can be
adjusted to a particular machine by an appropriate dimension
statement.
Various other essential parameters-such as the length of a stored
function, the number of stored functions, the length and number of
input-output blocks, the maximum number of simultaneously sounding
voices-will change with different computers and compositions. These
parameters have been assembled into the IP array, which is compiled
by a BL0CK DATA subprogram. Hence the parameters can be easily
changed.
The usual unit generators and GEN functions use fixed-point arith-
metic and store their results in the I array. (It would not be difficult to
use floating-point routines instead, or to use both.) However, the routines
do not produce F0RTRAN integers. Instead, F0RTRAN fixed-point
numbers are multiplied by 2n , which in effect puts their decimal points
n places from the right end of the memory words. Values of 2n for unit
generators and for GEN functions are also compiled into the IP array.
These values can be changed to accommodate different lengths of
memory word.
Output samples are written on a digital output tape by a combina-
tion of SAM0UT and FR0UT subroutines. These are inherently
machine-language operations, and there is no way to avoid so writing
them. However, they can be brief and demand little programming time.
Chapter 3, which presents the Music V manual, is organized in the
same manner as the program; it starts with a discussion of Pass I and
120 CHAPTER THREE

its subroutines and then proceeds to the other passes. The actual
F0RTRAN programs are, of course, the ultimate and best description
of Music V; they should be read along with the manual.

2. Description of Pass I
The purpose of Pass I is to read the input data (score) and translate
it into a form acceptable to the subsequent passes. The operation is
diagrammed in Fig. 49.

Fig. 49. Pass I.

statement

The interpretive input routine READ! (and READO, which is used

to read the first record) is written in F0RTRAN IV. It is designed for a
computer with a word length of 36 bits. It requires two user-supplied
subroutines (M0VL, M0VR) to be written in machine language for
purposes of character shifting. Minor modifications to READO and
 MUSIC V MANUAL 121

READ I are necessary for computers of different word length and for
different modes of input (see Section 7 for details).
The input data comprise a series of data statements punched in free
format in columns I through 72 of cards. A data statement need not
correspond to a single card.
A data statement begins with an operation code and is terminated by
a semicolon. Other fields of information in the statement are separated
by blanks (any number) or commas. Null fields, i.e., those denoted by
successive commas, are assumed to have the value O. With the exception
of statements used in instrument definitions (see Section 4), the fields
of a data statement are referred to as P fields since they load sequentially
into the P array located in C0MM0N storage in Pass I. 2 The operation
code, written as a three-letter mnemonic (see Section 3) is converted to a
numerical equivalent and goes into pel); the second field, containing
an action time that specifies when the operation corresponding to the
code is to be performed, goes into P(2). The other fields are interpreted
according to the specifications of the various operation codes. If a
field other than the OP code is written as an asterisk (*), the value
stored in the corresponding position of the P array will be the value
previously stored there. This feature can be employed to advantage
when parameters remain constant over a sequence of data statements.
The input data are terminated with the data statement having the
operation code of TER. Failure to provide this statement will result in
an error comment.
The input program makes certain checks on the data statements and
when errors are detected the value of IP(2), located in C0MM0N
storage, is set to 1. Since this location is initially 0, Pass I can verify at
its conclusion whether or not errors have been detected and, if so, the
run is terminated without proceeding to Pass II. Termination is accom-
plished by calling a nonexistent subroutine named HARVEY.
As the data cards are read, they are printed, and any error comments
are printed out after the offending statement. Data statements begin-
ning with operation code C0M result only in printing and are not
processed further. Such statements may be used to annotate the input
data with comments.
In addition to establishing the appropriate values in the P fields,
READ 1 counts the number of P fields in the data statement and sets
IP(1) (in C0MM0N storage) to this count. Pass I is then a1)le to process
the data statement as is required by the operation code and to write
2 C0MM0N storage in Pass I is arranged according to the statement,
C0MM0N IP(10), P(100), D(2000)
122 CHAPTER THREE

out the translated statement as N, (P(I), I = 1, N), where N = IP(1),

to be read by Pass II.
Pass I contains a data array D(2000) which may be used for general
storage and may in particular contain data for the PLF subroutines.
SVI and SIA data statements load the D array. (SVI 0 10 100; would set
D(10) = 100.) The following D variables have special significance.
D(4) = Sampling rate
D(S) = Stereo-mono flag
D(S} = 1 for stereo
D(S) = 0 for mono

3. Operation Codes and Corresponding Data Statements

The operation codes are listed in the following table.

Numerical
Value Mnemonic Purpose

1 N0T Play note

2 INS Define instrument
3 GEN Generate function
4 SV3 Set variable in Pass III
5 SEC End section
6 TER Terminate piece
7 SVI Set variable in Pass I
8 SV2 Set variable in Pass II
9 PLF Execute subroutine in Pass I
10 PLS Execute subroutine in Pass II
11 S13 Set integer in Pass III
12 SIA Set integer in all passes
13 a C0M Print comment

a This code number is used only by READ 1. A data statement beginning with
C0M is printed but is not processed further.

Remarks
1. Only the first three characters of the operation code mnemonic
are scanned; thus a user may write N0TE, INSTRUMENT, GEN-
ERATE, SECTI0N, TERMINATE, or C0MMENT in place of the
three-letter codes if he prefers.
2. Integer-valued P fields may be written with or without decimal
points.
 MUSIC V MANUAL 123

3. Null fields, those denoted by successive commas, are assumed to

be O.
4. Fields specified as * are assumed to have the value previously
stored there. This feature provides continuation over a sequence of
data statements.
Description of Data Statements
Each statement begins with the mnemonic operation code (at least
three letters). The second field must contain the time at which the
operation is to be performed. Therefore the descriptions that follow the
specifications will begin with the third field. All statements are
terminated by a semicolon.
1. N0T-Play note
P(3) Number of instrument on which note is to be played
P(4) Duration of note (in beats)
P(5). .. As desired by instrument referred to in P(3)
2. INS-Define instrument
P(3) Number of instrument being defined
3. GEN-Generate a function
P(3) Number of generating subroutine (see Section 25)
P(4) Number of function to be generated
P(5). .. As required by generating subroutine
4. SV3-Set variable(s) in Pass III, starting with variable N
P(3) Number of first variable to set = N
P(4) Value of variable N
P( 5) Value of variable N + 1
P(6) ... (Number of variables to be set is automatically deter-
mined by the word count.)
5. SEC-End section and reset time scale to zero
6. TER-Terminate piece at specified time relative to last section
7. SVI-Set variable(s) in Pass I, starting with variable N
P(3) Number of first variable to set = N
P(4) Value of variable N
P(5) Value of variable N + 1
P(6) ... (Number of variables to be set is determined by the
word count.)
8. SV2-Set variable in Pass II
Fields are as in SVI
9. PLF-Execute subroutine in Pass I
P(3) Number of subroutine: 1, 2, 3, 4, or 5
P(4). .. As required by subroutine referred to in P(3)
124 CHAPTER THREE

10. PLS-Execute subroutine in Pass II

Fields are as in PLF
11. SI3-Set integer(s) in Pass III, starting with integer N
P(3) Number of first integer to be set = N
P(4) Value of integer N
P( 5) Value of integer N + 1
P(6) ... (Number of integers to be set is determined by the
word count.)
12. SIA-Set integer(s) in all passes
P(3) Number of first integer to be set = N
P(4) Value of integer N
P( 5) Value of integer N + 1
P(6) ... (Number of integers to be set is determined by the
word count.)

4. Definition of Instruments
An instrument definition begins with the data statement" INS t n ; "
where t specifies the time at which instrument n is to be defined.
Subsequent data statements indicate the unit generators used in the
instrument and their associated parameters. The data statement
" END;" terminates the definition.
The unit generators that are recognized by name (i.e., three-letter
mnemonic) by READI follow.

Type
Name Parameters Number Purpose

0UT II,0 ; 1 Monophonic output

0SC II, 12, 0, F, S ; 2 Oscillator
AD2 II, 12, 0 ; 3 Two-input adder
RAN II, 12, 0, S, T1, T2 ; 4 Random function generator
ENV 11, F, 0, A, SS, D, S ; 5 Envelope generator
STR II, 12, 0 ; 6 Stereophonic output
AD3 II, 12, 13, 0 ; 7 Three-input adder
AD4 11, 12, 13, 14, 0 ; 8 Four-input adder
MLT II, 12, 0 ; 9 Multiplier
FLT II, 12, 13, 0 ; 10 Filter
RAH 11, 12, 0, S, T ; 11 Random and hold function
generator
SET II; 102 Set new function

See Section 5 for a more complete description of the unit generators.

 MUSIC V MANUAL 125

Data statements that specify unit generators may begin with the
three-letter mnemonic name or with the type number. READ1 recog-
nizes the 12 types listed in the table above by name, 3 and makes a
check on the proper number of parameters. If, for example, four or
six parameters are listed for 0SC, which requires five parameters, an
error condition will result, causing the job to terminate at the conclusion
of Pass I after all input cards have been scanned. Since unit generators
may be labeled by type number as well as name, it is possible to add
units to the subroutines F0RSAM (coded in F0RTRAN IV) or
SAMGEN (coded in basic machine language) used in Pass III without
the need for modifying READ1. Data statements referring to these new
units by type number will be accepted by READ 1, but no check will be
made for proper number of parameters.
The notation for these parameters used on the data statement is as
follows:
Pn refers to nth P field on note card
Vn refers to nth location in variable storage of Pass III
Fn refers to nth stored function
Bn refers to nth 1-0 block used by units
For example, instrument No.3 would be defined at t = 10 by the
following data statements:
INS 10 3 ;
0SC P5 P6 B2 F1 P30 ;
AD2 P7 VI B3 ;
0SC B2 B3 B2 F2 P29 ;
0UT B2 B1;
END;
READ 1 translates each mnemonic data statement into an all-numerical
data statement as follows:
(1) In all data statements, PI contains 2, the numerical equivalent
of INS, and P2 contains the action time (10 in the example).
(2) P3 contains the instrument number (3) in the first data statement.
(3) In the second through the last data statements, P3-Pn contains
the numerical equivalent of the mnemonic data statement fields
Pl ... P1ast , respectively. The name equivalents for the unit
generators are their type numbers listed above. The equivalents
of the P's, V's, etc., are as follows:
3 The "named" generators change frequently. The table describes the state of
affairs in April 1968 at Bell Laboratories.
126 CHAPTER THREE

Pm~m 1 ::;; m :::;; 100

Vn~100 +n
Fp ~ -(100 + p)
Bq ~ - q 1 ::;; q ::;; 100
The equivalents are unique because only 100 P's and 100 B's are
allowed. P's are represented by positive numbers from 1 to 100,
V's by positive numbers greater than 100, B's by negative numbers
from -1 to -100, and F's by negative numbers from -101 to
-00.
(4) The last mnemonic data statement, END, has only two fields,
PI = 2 and P2 = action time. It is recognized in Pass III by its
word count of 2; this terminates the instrument definition.
The example is translated into the following numerical data
statements:
2 10 3 ;
2 10 2 5 6 - 2 - 101 30 ;
2 10 3 7 101 -3 ;
2 10 2 - 2 - 3 -:.- 2 - 102 29 ;
2 10 1 -2 -1 ;
2 10 ;
All passes of the program operate exclusively on the numerical state-
ments; all mnemonics are translated by READ 1.

5. Unit Generators
fJUT: Output Unit (Numerical equivalent = 1)
Diagram:
I

Data statement: 0UT, I, 0 ;

6
Operation: This unit generator adds the specified input into the specified
output block thus combining it with any other instrument that con-
currently uses the output block. 0 1 = 0 1 + 11 where i denotes the ith
sample.
 MUSIC V MANUAL 127

Example: One of the simplest P5 P6

instruments is defined as
INS, 0,1;
0SC,P5,P6,B2,FI,P20;
0UT, B2, BI;
END;

BI is often used as the output block. The location of the output block
must be compiled into IP(lO) (see Section 17).

@SC: Oscillator (Numerical equivalent = 2)

Diagram: I I 12

Data statement: 0SC, II, 12, 0, Fj, S;

¥ f/)

Operation: The oscillator generates functions and oscillations according

to
0 i = IIi· Fj([Sd mod function length in samples)
and
Si+ 1 = Si + 121
So = initial value of sum
where 0 1 is output, III is amplitude, Fj is a (stored) function, Sl is the
sum, 121 (increment) determines the frequency of oscillation, and i
indexes the samples.
The frequency of the oscillation is
_ sampling rate x 12
F requency - f . Iength In
. sampIes
unctIOn
The length of the function in samples is equal to IP( 6) - 1. n
(which = IP(6)) samples of each function are stored. The first and
nth samples represent the same point on the function and must have
128 CHAPTER THREE

the same value. Hence the function is periodic with period n - I

sample times. One note parameter P n must be reserved for the sum.
The value of this parameter on the data statement determines the
initial value of the sum So. Usually n is selected to be one of the last
locations in note parameter storage; if Pn is not written on the note
card (cf. Section 4) Pn is automatically set to zero at the beginning of
each note.
Example: The example for the output box is also appropriate for the
oscillator. FI determines the wave shape. P5 is the amplitude. P6
determines the frequency. Specifically
_ P6 x sampling rate
F requency - f . Iength In
. sampIes
unctIOn
See Chapter 2, section on 0SC Generator, and Chapter 3, Section 6,
for more details about 0SC.
AD2, AD3, AD4: Two-, Three-, and Four-Input Adders (Numerical
equivalent = 3; AD3 = 7; AD4 = 8)
Diagram: I I I2

Data statement: AD2, I I, 12, 0 ;

Operation: Output is generated according to
0 1 = 111 + 121
The other adding units (AD3 and AD4) work in a manner analogous
to AD2.
Example: None.

RAN: Random Function Generator (Numerical equivalent = 4)

Diagram: II 12

RAN
 MUSIC V MANUAL 129

Data statement: RAN, 11, 12, 0, S, Tl, T2;

Operation: Output is generated according to
0 1 = 111* R i(I21)
RAN generates a low-pass random function whose peak amplitude
is 111 and whose cutoff frequency is controlled by 121 and is approxi-
mately
B f'V sampling rate. 121
f'V 2 512
More specifically, Rl is a function, varying from -1 to + I, obtained
by sampling the line segments that connect independent random
numbers, N 1• There are 512/12 samples between each pair of independent
random numbers (see Fig. 50). The Nt's are uniformly distributed from
-1 to + 1.

+1

o ~~-~-----::l~t-------+---'--- Samples

-I

etc.

Fig. 50. Random function.

S, Tl, and T2 are temporary storage locations which are normally

kept in note-parameter locations. S holds a sum equivalent to the 0SC
sum. Tl holds N i - 1 and T2 holds Nl - N i - 1 where N i - 1 and Ni are
the last two independent random numbers.
Example: A typical instrument to P7
produce a band-pass noise:
INS, 0, 1;
RAN, P5, P6, B2, P30, P28, P27 ;
0SC, B2,P7,B2, Fl,P29;
0UT, B2, BI;
END;
130 CHAPTER THREE

Function F1 is assumed to be a sine wave. By means of the modula-

tion inherent in the multiplication of the left oscillator input, B2 will
contain samples of a band-pass noise whose center frequency is
P7 x sampling rate
function length
and whose bandwidth is
P6 x sampling rate
512
The peak amplitude is P5.

ENV: Envelope Generator (Numerical equivalent = 5)

Diagram: II 12 13 14

Data statement: ENV, II, F j , 0, 12, 13, 14, S;

Operation: This unit scans a function F j at a variable rate to produce an
attack, steady-state, and decay amplitude envelope on a note.

0 i = IIi * F j (scanned according to 12, 13, and 14)

The first quarter of F j gives the attack shape, the second quarter of F j
gives the steady state, the third quarter of F j gives the decay shape, the
fourth quarter is unused and should be zero.
Specifically, the sections of F j and the scanning rates are shown in
Fig. 51.

Scanning rates:

~.
I2lacations/somple
13 loc.lsam. 14 loc.lsam.
~I~I~I
I I I
I I I
I I :
I I I

i : :
At tack ~ Steody state-':_ Decoy I

o 1/4 FL 1/2 FL 3/4 FL FL

Fig. 51. Envelope function. FL = function length in samples.

 MUSIC V MANUAL 131

In a typical use

12 = function length in samples

4· attack time· sampling rate
13 = function length in samples
4· steady-state time· sampling rate
14 = function length in samples
4· decay time· sampling rate

S is a temporary storage location (note parameter) to store a sum

similar to the sum in 0SC.
Example: The principal use is P5 P6 P7 P8 P9
to generate envelopes
for notes.
INS, 0, 1;
ENV, P5, FI, B2, P6, P7, P8, P20;
0SC,B2,P9,B2,F2,PI9;
0UT, B2, BI;
END;

P6, P7, and P8 determine attack, steady state, and decay times, respec-
tively. P5 determines the maximum amplitude. P9 determines the
frequency. FI determines the envelope and F2 the oscillator waveshape.
Typically P6, P7, and P8 are computed by an elaborate C0NVT
function (see Chapter 2, section on Additional Unit Generators, ENV).

STR: Stereophonic Output Box (Numerical equivalent = 6)

Diagram:

Data statement: STR, 11, 12, 0;

132 CHAPTER THREE

Operation: This unit puts alternate samples from II and 12 into 0

0 21 = IIi
0 2 1+1 = I2i
This arrangement is suitable for a stereophonic output conversion.
The stereophonic output requires an output block length equal to
two input-output block lengths. Typically Bl and B2 are set aside for
output storage.
Example: Two instruments are defined P5 P6 P5 P6
which are identical except
that one uses the left
channel and the other the
right.
INS, 0, I;
Vi
0SC, P5, P6, B3, FI, P20;
STR, B3, VI, BI ;
END;
INS, 0, 2;
0SC,P5,P6,B3,FI, P20;
STR, VI, B3, BI ;
END;
VI is assumed to be zero. Note that blocks BI and B2 have been reserved
for output.
In another example, a single instrument produces sound in both
right and left channels.
P5 P6 P7 P8 INS,O, I;
0SC,P5,P6,B3,FI,P20;
0SC,P7,P8,B4, F2,P21;
STR, B3, B4, BI ;
END;

RAH: Random and Hold Function Generator (Numerical equivalent =

11)
Diagram:
 MUSIC V MANUAL 133

Data statement: RAH, 11, 12, 0, S, T;

Operation: Output is generated according to
0 i = IIi X Rn(I2i)
where Rn(I2i) is a succession of independent random numbers which
change every 512/12 samples. Thus this generator holds each random
number for 512/12 samples. Rn(12i) is uniformly distributed from -1
to + 1.

+1
..... .....
RI
R2

• • ~3• •
... ..
-1
5ii samples

Sand T are temporary storage locations which are normally kept in

note parameter locations. S holds a sum equivalent to the 0SC sum·
T holds the current'Ri •
Example: A typical instrument P5 P6 P7 P8
to produce a succession
of random pitches:
INS, 0, 1 ;
RAH, P7, P8, B2, P20, P19 ;
AD2, P6, B2, B2 ;
0SC,P5, B2,B2, F1,P18;
0UT, B2, B1;
END;

Function F1 can be any desired waveform. P7 should be at most equal

to P6. The pitch frequency will assume a succession of random values
between the frequencies
(P6 - P7) x sampling rate
function length
and
134 CHAPTER THREE

(P6 + P7) x sampling rate

function length
A new value of the pitch frequency is generated every 512/P8 samples.

SET: Set New Function Number in Unit Generator (Numerical

equivalent = 102)
Diagram:
I I

Data statement: SET, II ;

~
Operation: SET enables changing the function number in an 0SC or
ENV unit generator by specifying the new function number as a note
parameter.
In the instrument definition, SET must be just ahead of the unit
generator on which it is to act; the input specifies in which P field of the
note card the new function number is to appear. If this P field is given a
negative or zero value, no change is effected; if it is given a positive
integer value, this value is the new function number.
Example: INS, 0, 1 ;
SET,P7;
0SC,P5,P6,B2,Fl,P20;
0UT, B2, Bl;
END;
With this instrument definition, all three of the following note cards
N0T, 0, 1, 1, 1000, 50, 0;
N0T, 1, 1, 1, 1000, 50, 1 ;
N0T, 2, 1, 1, 1000,50, -2;
will leave function #1 in 0SC, whereas the note card
N0T, 3, 1, 1, 1000, 50, 2;
replaces function #1 by function #2 in 0SC.

6. Special Discussion of esc Unit Generator 4

Probably the most basic and important unit generator used by
Music V is the oscillator. Since the oscillator utilizes most of the basic
4 This discussion of 0SC was provided by S. C. Johnson.
 MUSIC V MANUAL 135

principles of Music V, a detailed description of its operation should

prove useful in the design and implementation of additional unit
generators.
The oscillator is a unit generator, meaning that it is a "device" that
is useful in building "instruments." This device is simulated by a
general computational algorithm which can produce any periodic
function at any frequency or amplitude. This algorithm should be
quite efficient, since it must compute 10,000-20,000 numbers for each
second of sound.
Efficiency and generality are gained through the use of stored
functions. The values of a stored function need be computed only once
(by a GEN subroutine in Music V) and then may be referred to by any
unit generator. By making the functions interchangeable among unit
generators, we need keep only one copy of any function used and one
copy of any unit generator in the computer memory.
The mathematical algorithm for simulating an oscillator is described
by the equation

and
Si+l = S1 + Ii
where
01 = the ith output sample
Ai = the ith amplitude input
Ii = the ith increment input (controls frequency)
F = a stored function (controls waveshape)
Si = the ith sum of increments
FL = the length of the stored function (in samples)
Assume for a moment that the stored function is a representation of
a sine wave occupying 101 computer locations, F(O), F(2), ... , F(100) .

+1
•••••• •
• ••
••• ••
•
O~~----------~·'--------------.~I~--
• • I
• • I
• • I

-I
etc
•
•
.... . •
• I

:
I

FUOO)
136 CHAPTER THREE

The value of F(O) is sin (0/100 * 27T), F(1) is sin (1/100 * 27T), F(2) is
sin (2/100 * 27T), etc. Since °: :;
Isin xl :::; 1.0, we may multiply the
values of the function by any amplitude A to produce output samples
in the desired range, 0 :::; 1011:::; A.
How does the oscillator reproduce this sine wave at any frequency?
Assume that we have fixed the sampling rate at 10,000 samples per
second. This means that the digital-to-analog converter will convert
10,000 samples into sound every second, and each sample number we
output represents 1/10,000 second of sound. If we multiply the stored
function shown above by an appropriate amplitude and output it
directly, then each period of the wave will contain 100 samples, and
it will be heard 10,000flOO or 100 times per second. This corresponds
to a frequency of 100 Hz. Since the sampling rate is fixed, to double the
frequency of the sound we must halve the number of samples per period
of the wave. We do this simply by referring to every other value of the
stored function.

,
• '
.: ,l .. ••
I •

• I I I •
, I I I •
·1 I I I ••
• 1 ~I I I. I

1(=2)
• •
•
•
•• ••
•
••• •
Thus the output samples will be given by the relations
0(1) = F(O) * A (s = 0)
0(2) = F(2) * A (s = 2)
0(3) = F(4) * A (s = 4)

etc.
The output wave then has 50 samples per period and is heard at
10,000/50 or 200 Hz. To obtain the output 0 1 in this case, the inde-
pendent variable in the function F(s) is incremented by 2 each time the
function is referred to. If the increment used was 4, we would output
100/4 = 25 samples per period, or a sine wave of 10,000/25 or 400 Hz.
In general then
. sampling rate
Frequency III hertz = I . d
samp es per peno
and
 MUSIC V MANUAL 137

S 1 . d function length
amp es per peno = increment

therefore,

. h t _ sampling rate * increment

F requency III er z - f . 1
unctIOn ength

and

. 1) function length * frequency in hertz

I ncrement (III samp es = 1.
samp Illg rate

Modulus arithmetic is used in conjunction with the cumulative sum

of increments S1 in order to achieve periodicity in the references to the
stored function.
A final point concerns the sum of increments. Assuming a function
length and sampling rate as above, the increment necessary to produce
a 150-Hz tone I is (100 * 150)/10,000 or 1.5. Obviously any continuous
function will have a value at S = 1.5, but we cannot directly talk about
the I.5th computer location of stored function F. Three approaches to
this problem have been used: truncation, rounding, and interpolation.
The fastest method is truncation, where the greatest integer [S] con-
tained in the sum of increments is used as the S value. This is easily
accomplished with fixed-point computer arithmetic, but may lead to
some distortion of the output (see the table below). In the rounding
method, we round the sum of increments to the nearest integer and use
this as the S value. Although this takes a little more computation, it
leads to better results.
In the interpolation method, the sum of increments is truncated to
obtain a function value as in the truncation method. This function
value is then corrected by linear interpolation: if y is the function value
at F([S]), y' is the function value at F([S] + 1), and h is the amount by
which the sum of increments exceeds [S] (= s - [S], or the fractional
part), then the corrected value of the function is y + (y' - y)h. This
method takes the most computer time but in practice produces the
greatest accuracy. It can also effect a saving of memory space in the
computer, since, as is shown in the table below, treating a stored
function of 512 locations with truncation produces a greater distortion
of the output than using interpolation on a function only 32 locations
long.
138 CHAPTER THREE

The table shows the results of computing 500 values of sine x, using
various methods and stored function lengths. The table entries are the
percentage rms error.

Function Length Truncation Rounding Interpolation

32 7.9 4.0 0.3

64 3.8 1.9 0.06
128 1.7 1.1 0.02
256 0.9 0.5 0.004
512 0.5 0.2 0.001
1024 0.24 0.12 0.0002

In general, rounding is about twice as accurate as truncation, and

doubling the length of the stored function doubles the accuracy for
both the truncation and rounding. Doubling the function length
quadruples the accuracy for the interpolation method, however. Which
method is used will depend on the availability of computer time versus
memory space in a particular installation of Music V.
The distortion level of the oscillator depends on the function length
and the particular numeral process used. It also depends on the
particular increment used: distortion occurs only when the increment
is not an integer. Finally, it depends on the wave shape used: the dis-
tortion level will increase when the slope of the stored function is steep
at the point considered.
How much this distortion alters the quality of the sound is hard to
predict; a function with steep slope should be expected to be more
distorted than a sine wave, and yet in many cases the distortion will be
more audible with sine waves than with complex waveforms. For
instance, the synthesis of a frequency-modulated sine wave with the
following parameters:
Function length = 512 samples
Sampling rate = 10,000 Hz
Frequency deviation = 3 % (of fundamental frequency)
Vibrato rate = 25/0 (of fundamental frequency)
produces a clearly distorted sound when truncation is used, and an
acceptable sound when interpolation is used. But if the sine wave is
replaced by a complex tone with harmonies decreasing at 6 dB or 12 dB
per octave, there is almost no audible difference between sounds
synthesized with truncation and those synthesized with interpolation.
 MUSIC V MANUAL 139

7. Input-Output Routines for Pass I and Pass II

Input/or Pass I: READO and READ1
The interpretative input routine for Pass I is a F0RTRAN IV sub-
routine named READI. It has an additional entry point called READO
used for reading the first record. The program, as supplied with Music V,
is designed for a 36-bit word machine and accepts input data punched in
the free format in columns 1 through 72 of cards, as has been described
in Section 3.
READO reads an initial record into the input buffer ICAR (equivalent
to CARD). The characters are stored one per computer word and are
shifted to the right end of the word by the M0VR subroutine. (M0VR
is one of the machine-language routines necessary for Music V.)
The operation of READ1, the main program, is diagrammed in
Fig. 52. After writing out a record to set Music V to stereo or mono
(which will be discussed below), the program (at 10)-> scans to the end
of the first data statement marked by ";". If necessary, more input
records are read.
The characters are organized (at 21) into fields with exactly one blank
character separating successive fields. The organized data are stored in
IBCD and are printed out. The first field is compared with all possible
mnemonics that may be written in it. If a match is found, the numerical
equivalent of the mnemonic is found and one of a number of branches
(at 29) is taken depending on the value of the first field.
If no match is found for the first field (at 40), it is taken to be a
number if the data statement is inside an instrument definition. Other-
wise, an error comment is made and the statement is rejected.
The remaining fields on the data statement are converted to numerical
form by one of several sections of the program (218, 201-210, 100,
300-1200, 200, 217, 220, and 30) depending on what the first field is
and whether the data statement is part of an instrument definition.
All score records have a mnemonic operation code as the first field
and an action time as a second field except the second through the last
cards in an instrument definition. In an instrument definition, such as
the one given below
INS, Action time, Inst No;
0SC,P5, P6, B2, FI,P20;
0UT, B2, BI;
END:
5 The numbers cited in the descriptions of programs refer to statement numbers
in the F0RTRAN program. These numbers are also shown in the block diagrams.
 Premature
E.O.F., set
error flag
and write
TER dota
siatement

Organize statement into fields

store in IBCD, print statement
determine numerical equivalent
of first fie I d

0UT N0T INS

fl)SC GEN END
AD2 SV3
RAN SEC
ENV TER
STR SVI
AD3 SV2
201-210
r---"'--,

yes
Convert BCD
field from
L pointer
to next
bla nk to
floating
point
number

Fig. 52. Block diagram of READl

Return on computed
program. G0 T0 n NI
 MUSIC V MANUAL 141

the operation code and action time appear only on the first data record.
READI takes the operation code (= 2) and the action time from the
first data record and stores these in P(1) and P(2) in all subsequent
records connected with instrument definition. The value of P(3) is the
type number of the unit, and the remaining fields are interpreted and
converted to floating point and stored starting in P(4). Word (i.e., field)
count for the statement is established in IP(I).
The conversion from BCD to floating point is done by a subroutine
(at 70) which finds the position of the decimal point in the field of
characters (or supplies it at the end if missing) and then multiplies the
characters, which are expressed as integers, by the appropriate power of
10 and sums over all characters in the field.
Any errors that are detected cause an error comment to be printed
below the printout of the data statement in which the error occurred.
In cases other than that of an incorrect operation code, the entire
statement is scanned so that all errors will be detected. Incorrect
operation codes, however, prevent proper interpretation of the remain-
ing fields in the data statement. When errors occur, a flag is set in
C0MM0N storage (namely IP(2) is set to 1) so that Pass I may ter-
minate the job at its conclusion. Furthermore, when errors are de-
tected, the data statement is not returned to Pass I but control returns
to the entry point of READ 1 to obtain the next data statement.
It will be noted that the input array for the card data is named CARD
which is (F0RTRAN) equivalent to ICAR. Also, IBCD is equivalent
to BCD. This equivalence is necessary because the characters when read
in with format Al require a floating-point designation. However, for
purposes of comparison, the data must be regarded as in integer form.
Hence, the characters must be right adjusted (moved to the right and
of the computer word). Similarly, when the organized data statement is
to be printed out, it must be put back into left-adjusted form so that it
may be printed out in Al format. Consequently, this routine uses two
subroutines, which must be written in machine language and, therefore,
supplied by the user. READI (and REA DO) makes the following calls:

CALL M0VR (CARD, NC)

CALL M0VL (CARD, NC)

The characters stored in NC consecutive locations of CARD are

right (left) adjusted and replaced in the same locations. Calls to M0VR
are found after the two "read" statements, and the" print" statement
is preceded by a call to M0VL and followed by a call to M0VR.
142 CHAPTER THREE

READ 1 inspects the output unit generators (0UT or STR) in the

instruments, and if a change from monophonic to stereophonic opera-
tion or vice versa is made, it writes out an appropriate stereo or mono
control record at the end of the instrument definition. The record is

SIA TA 8 (01 (= mono») .

(= stereo) ,

The action time TA is the same as the action time for the instrument
definition. Inspection is done by the first statement in READ 1. END
equals 1 at the end of an instrument definition and equals 0 otherwise.
SNA8 equals 1 if the mono-stereo mode is changed and equals 0
otherwise. STER = 0 if the last out box is 0UT, and STER = 1 if
the box is STR. Music V is assumed initially to be in the monaural
mode.
If the program is to be run on a machine of different word length,
the F0RTRAN DATA statements for arrays IBC, IVT, and L0P
must be changed. These contain right-adjusted BCD characters for the
break characters used in delimiting the input, the parameter types P,
V, F, and B used in specifying unit generators, and the characters used
in the three-letter mnemonic names for operation codes. In a 36-bit
machine such data are entered as 6HOOOOOX, in a 24-bit machine as
4HOOOX. If the input data are to be read in from any medium other
than cards, the two "READ" statements at 15 and under READO
need to be changed as required. The number of characters obtained by
executing a "READ" command is a variable NC, established in this
version of the program as 72. The arrays that hold these data have been
dimensioned to accept a maximum value of NC equal to 128.
The break characters delimiting the fields of input data are the
blank, comma, and semicolon; NBC, the number of break characters,
is equal to 3. (If typewriter input is to be substituted, an additional
break character may be needed equal to the carriage return.)
The most frequent change in READ 1 is the addition of other OP
codes or unit-generator names. The following steps will accomplish
this change:

(1) Add the three-character mnemonic to the end of the L0P array.
(The size of the L0P array mayor may not need to be increased,
the word count on the L0P DATA statement must be increased.)
(2) Increase N0PS by 1.
(3) Put another branch at the end of the G0 T0 at 29.
 MUSIC V MANUAL 143

(4) Write appropriate code for the branch. The code for branches
201-210 or 300-1200 will usually serve as a model.
A few of the variables in the program are
NUMU Normally equals O. It is set to 1 to disable checking the
number of fields in a unit generator.
NPW The number of fields expected in a unit generator, not
counting the name. For example, AD3 PI P2 P3 B1 ;
would have NPW = 4.
L Scanning index for the IBCD array. It normally points
to the character just ahead of the next field to be pro-
cessed. L is changed by many parts of the program,
including the BCD to floating-point converter.
I Scanning index for the ICAR array.
J Scanning index to store characters in the IBCD array.

Output for Pass I: WRITE]

The number of fields in a data statement is established by READ 1
and stored in IP(l) located in C0MM0N. The fields of the data state-
ment have been interpreted and the P array appropriately filled by
READ 1. Thus after Pass I has properly processed the data statement,
it may call WRITE1 to write the data statement on tape or disc so as to
be available to Pass II. The call from Pass I is as follows:
CALL WRITE 1 (10)
WRITE1 sets N = IP(l) and writes the list N, (P(I), I = 1, N) onto
data file 10 in binary format.

Input for Pass II: READ2

Pass II calls upon subroutine READ2 with the call CALL READ2
(10) to read N, (P(I), I = 1, N) from data file 10 and establish
IP(l) = N.

Debug READO and READ]

For testing purposes, an all-F0RTRAN score-reading program is
provided to replace READO and READ 1. The program reads an all-
numerical score. Each data statement has the following information:
N P(l) P(2) ... peN)
where N is the word count and the .subsequent fields destined for the
first N locations of the P array.
144 CHAPTER THREE

The calling sequences are the same as those of READO and READ I,
namely
CALL REA DO
and
CALL READI
Debug READO does nothing.
Debug READ 1 reads one statement into the P array according to the
F0RTRAN statement
READI, K, (P(J), J = I, K)
where the format statement 1 is
1 F0RMAT(l6, IIF6.0/(12F6.0))
Thus twelve numbers are read from the first 72 columns of each card.

8. PLF Subroutines
A data statement of the type
PLF 0 n D4 D5 ... Dm ;
will cause the following call to take place during Pass I at the time the
data statement is read
CALL PLFn
where n is some integer between 1 and 5. PLFn is a subroutine which
must be supplied by the user. These subroutines can perform any
function desired by the user. Usually they will generate data statements
for Pass II or manipulate Pass I memory (the D(2000) array).
The information of the data statement PLF, 0, n, D4, ... , Dm will be
placed in the P(100) array in P(1) - P(m) at the time PLFn is called.
The P, D, and IP arrays are kept in common storage and hence are
available to the PLFn routine. The dimension and common statements
in the PLF routine and in Pass I must, of course, agree. For examples
and a further discussion of PLF subroutines, see Chapter 2, section on
Composing Subroutines-PLF.

9. General Error Subroutine

A general-purpose ERR0R subroutine is used by all three passes. A
statement
CALL ERR0R(N)
 MUSIC V MANUAL 145

will cause the following comment to be printed

ERR0R 0F TYPE N
where N is an integer.
The meaning of N is as follows
N

Pass I errors
Nonexistent OP code on data statement 10
Nonexistent PLF subroutine called 11
(i.e., in call PLFn , n < 1, or n > 5)
Pass II errors
Too many notes in section, D or I array full 20
Incorrect OP code in Pass II 21
Incorrect OP code in Pass II 22
Nonexistent PLS subroutine called 23
(i.e., in call PLS n , n < 1, or n > 5)
Pass III errors
Incorrect OP code in Pass III
Code is < 1 or > 12
Too many voices simultaneously playing a 2
Too many voices simultaneously playing 3

a The maximum number of voices must be equal to or less than the number of
note parameter blocks (see Section 16).

In addition to these error comments, the READ 1 will print error

comments if it detects errors in or cannot interpret any data statements.
These comments are described in Section 7.

10. Description of Pass II

Pass II performs three general functions:
(1) Sorting the data records obtained from Pass I into forward
chronological order according to starting times,
(2) Applying special conversions to some of the input data records
by calling the user-supplied C0NVT subroutine, and
(3) Applying the metronomic time-scaling operations to starting
times and durations.
After completing these functions, Pass II writes its output onto data
file 11 for subsequent use by Pass III (if desired, a Pass II report is
printed, see WRITE2 below). The entire operation may be diagrammed
as shown in Fig. 53.
146 CHAPTER THREE

Initialize
section

Read section
of data CALL READ2

CALL S0RT

CALL WRITE2 ---+ CALL CONVT

no
Fig. 53. Pass II block diagram
-Music v.

Pass II maintains the following arrays in unlabeled common storage

C0MM0N IP(IO), P(lOO), G(IOOO), 1(1000), T(lOOO), D(IO,OOO)
After setting certain variables (standard sampling rate, size of D and 1
arrays, and number of OP codes) to their initial values, Pass II calls on
READ2(lO). READ2 then reads information from data file 10 accord-
ing to the format: (K, P(I), I = I, K). The value of K is stored into
IP(l). This call is repeated until an entire section has been read in and
the data statements are accumulated in the D array. The 1 array is used
to hold subscripts that point to the beginning of each data record in the
D array. The T array is used to hold the action time of each data
statement P(2). After an entire section has been read in, Pass II sorts the
T array into ascending numerical order by calling S0RTFL and S0RT
(S0RTFL is merely an initialization routine which informs the sort-
program package that floating-point numbers are about to be sorted).
 MUSIC V MANUAL 147

It also sorts the I array (pointers) as a passive list on T, so that after

S0RT has been called this point list has been rearranged according to
the starting times of the data statements.
Each data statement is then accessed from the D array according to
the order specified by the I pointers (chronological order of action
times). Each data statement is inspected to see:

(1) If the OP code is 8 or 12 (SV2 and SIA, respectively), then the

variable list (i.e., P(4) through P(4 + n)) is stored into the G
array starting at G (P(3)).
(2) If the OP code is 10 (PLS), then a call to PLSn is generated, where
n is the number stored in P(3).
(3) If the OP code is 7 or 9 (SVI and PLF, respectively), an error
message is printed (error of type 22). This error is not fatal,
however, and Pass II merely ignores the offending instruction.
(4) If the OP code is 1, 2, 3, 4, 5, 6, 11, 12 (N0T, INS, GEN, SV3,
SEC, TER, SI3, SIA), the subroutine WRITE2(ll) is called.
This subroutine applies the optional metronome time-scaling
operations, prints the optional Pass II report, calls the subroutine
C0NVT to modify any note parameters, and writes out a record
on data file 11 for subsequent use as input to Pass III.
The record is written according to statements
K = IP(l)
WRITE(11), K, (P(J), J = 1, K)
K, which is kept in IP( 1), is the word count. Data are in the
P(lOO) array. Details of the operations done by WRITE2 are
discussed in Section 11.

After a section has been processed, the next section is read. The
section-reading sequence is terminated by a TER card via a flag lEND
which is set to 1 when TER is encountered. lEND is checked after each
section is processed.
The error comments produced by Pass II are printed by ERR0R and
are discussed in Section 9.
Pass II contains a general-purpose memory, G(lOOO), which is
primarily used by the PLS subroutines and by the C0NVT function.
Blocks of locations starting at G(n) may be set with a SV2 AT n x ... ;
record. The setting occurs at the action time, AT, relative to the other
data records.
Certain locations in the G array have special functions:
148 CHAPTER THREE

G(1) Flag controlling Pass II report (= 0, print report; = 1,

suppress report)
°
G(2) Time-scaling flag (G(2) = for no time scaling; G(2) = n
for time scaling where metronome function starts at G(n))
G(4) Sampling rate
G(5) Starting beat of note
G(6) Duration of note in beats
G(8) Stereo-mono flag (= 0, mono; = 1, stereo)
The IP array contains certain other parameters:
IP(I) Word count for current record in P array
IP(2) Location in D array of beginning of data statement that is
currently calling a PLS subroutine
IP(3) Number of data statements in the D array

11. WRITE2
Pass II calls WRITE2(1l) in order to:
(1) Invoke the optional metronome operations described below
(2) Produce the optional Pass II report on the printer
(3) Call C0NVT to modify data record parameters
(4) Write (N, P(I), J = 1, N) on data file 11 for subsequent use by
Pass III
In order to utilize the metronome operations available in Pass II, a
nonzero value must be stored in the array location G(2). This value is
the beginning subscript in the G array of a tempo function such as the
one shown in Fig. 54. This is a function constructed of any number of
240
-5 180 M (=160)
c
.~ ~
.= 120
4
M (=126)
5
~~ 6 Mo(=60)
I-~
O~ __ ~~ ____________ ~ ____ ~~~

Bo BI B2
(=0) (=10) (=15)
Time (In beats) -+

Fig. 54. Tempo function.

line segments. It is stored, beginning at G(G(2)), as an arbitrary-length

list of number pairs, Bo (= 0), M o, B1 , M 1 , B2 , M 2 , ••. , Bm Mn where
Mi is the standard metronome marking (in beats per minute) at beat
 MUSIC V MANUAL 149

Bi of the composition. See Chapter 2, section on Compositional

Functions, for additional discussion.
WRITE2 uses the F0RTRAN function C0N(G, I, T) (see Section
12) to calculate the value at beat T of the function which is stored at
G(I). WRITE2 then converts P(2), the starting time, and P(4), the
duration, 6 from beats to seconds according to the following two
relationships:

Ti = T i - 1 + (Bi - Bi - 1 )· (F~~i))

where
Ti is current time in seconds, which replaces the value in P(2)
Bi is current beat number, the value found in P(2)
F(Bi) is the value of the tempo function at beat Bi and
Ti - 1 and Bi - 1 are the time and beat of the previous data record
and

Di = Li · (F~~i))
where
Di is the duration in seconds
Ll is the duration of note in beats and
F(Bi ) is as above
The tempo function itself may be placed into the G array via an
SV2 instruction. The function shown in Fig. 54, for example, could be
placed in the G array beginning at G(30) by the data records:
SV2, 0, 2, 30 ;
SV2, 0, 30, 0, 60, 10, 60, 15, 120, 45, 40, 56, 160, 63, 126
These metronome operations can be turned off at any time by setting
G(2) at 0. If the metronome operations are so turned off, P(2) and P(4)
are not affected by WRITE2 and are assumed to be in seconds.
The Pass II report is printed automatically by WRITE2 if G(l) = 0.
The Pass II report may be suppressed by setting G(l) # with an SV2 °
instruction (e.g., SV2, 0, 1, 1 ;). It consists of each data statement
printed in order of ascending action times. Each data statement is
shown exactly as it is presented to Pass III (if the data statements do
6 Durations are given on N0T cards only. P(4) is affected if and only if P(1) = 1
(playa note).
150 CHAPTER THREE

not exceed 10 fields, they are printed one per line; longer data state-
ments are continued on next line). In addition, if the metronome func-
tion is in use, P(2) and P(4) will have been converted into seconds, and
the original values of these parameters (in beats) are printed to the
right of each print line.
WRITE2 calls C0NVT immediately before it returns. G(5) and G(6)
contain the original values of P(2) and P(4) if metronomic scaling was
used.

12. C0N-Function Evaluator for Line-Segment Functions

C0N evaluates functions formed from a sequence of line segments.
These functions are useful for representing compositional functions
such as the metronomic marking. C0N is used by the time-scaling
routines in Pass II. It may also be used by PLF and PLS subroutines.
C0N can be evoked by a statement such as
Y = C0N(G, I, T)
which will set Y equal to the value at time T of the function stored at
G(I).
C0N expects to find a pair list in the G array beginning at subscript
J. The form of the list is Xl, Y h X 2 , .•• , Xm Y n, where Xi and Yi are the
abscissa and ordinate values for the breakpoints of the function. As
many breakpoints as desired may be used. Breakpoints do not need to
be equally spaced along the abscissa. If T falls between two breakpoints
(as it usually does), C0N computes Y as a linear interpolation between
the adjacent breakpoints.
As an example, the function shown in Fig. 55 would be stored
starting at G(30) in Pass II by the statement
SV2 0 30 0 1 10 12 20 1 ;
Its value at 13 would be obtained by
Y = C0N(G, 30, 13) = 8.7

13. S0RT AND S0RTFL

S0RT and S0RTFL are two utility routines in the Bell Laboratories
utility library on the GE645 computer. They are called by Pass II when
it arranges the data statements in chronological order according to
action times.
 MUSIC V MANUAL 151

12

10

o 10 13 20

Fig. 55. Linear interpolation between two breakpoints.

S0RTFL is an initializing routine called to specify that floating-point

numbers are to be sorted. The calling sequence for S0RTFL used in
Pass II is simply
CALL S0RTFL
The calling sequence used in Pass II for the sort routine is
CALL S0RT (T(l), T(2), IN, I)
where T(l) and T(2) are the first two words on a list that will be sorted
into monotonic increasing order, IN is the number of words to be sorted,
and I is the location of the first word of a second list which will be
rearranged in exactly the same manner as the T list.
In Pass II, T contains action times for data statements and I contains
pointers to where the data statements begin in the D array. After sorting,
the pointers in the I array will have been so rearranged that successive
pointers point to data statements in the D array in their proper
chronological sequence.
If two entries in the T array are equal, S0RT will not interchange
their order. The preservation of order is essential for the data statements
that define an instrument. All these have the same action time, but their
order must be maintained.

14. PLS Routines

A data statement of the type
PLS AT n D4 D5 ... Dm ;
causes Pass II to execute the statement
CALL PLSn
152 CHAPTER THREE

where n is some integer between I and 5. The call is carried out at action
time AT relative to the processing of other data statements in Pass II.
PLSn is a subroutine that must be supplied by the user. It can perform
any desired function, but a typical use would be to change a note
parameter, such as pitch, according to some composing rule. For more
information on the use of PLS routines, see the tutorial examples in
Chapter 2.
The data statement PLS AT n D4 . . . is stored in numerical form in
the Pass II D(10,000) array at the time the call to PLSn takes place.
The arrangement is
D(M) = word count
D(M + I) = 10 (the numerical equivalent of PLS)
D(M + 2) = AT
etc.
M = IP(2)

Thus, for example, in order to find D5 from the data statement, PLS
must look up M at IP(2) and then look in G(M + 6). Such a roundabout
procedure is necessary because of the sorting.
The dimension and common statements in Pass II and PLSn must,
of course, be identical.

15. C0NVT-Convert Subroutine

This subroutine is caned by WRITE2 immediately before each data
statement is written out to be used as input by Pass III. It must be
supplied by the user and replaces CVT functions in Music IV. Special
conversion of input parameters are possible, such as converting a
frequency given in cycles per second to an appropriate increment,
conversion of a special amplitude notation to a form acceptable to
Pass III, and so forth. Attack, steady-state, and decay times may be
converted to correct increments for driving the ENV generator.
The necessary F0RTRAN C0MM0N statement is
C0MM0N IP(IO), P(lOO), G(IOOO)
C0NVT is called by the statement
CALL C0NVT
At this time the parameters for the data statement are in the P array,
and the number of parameters is in IP(1). G(5) and G(6) contain the
 MUSIC V MANUAL 153

starting time and duration in beats, if the metronomic scaling has been
used.
C0NVT may perform complicated logical functions. It may increase
or decrease the number of parameters, changing IP(1) accordingly.
For more information, see the tutorial examples described in Chapter 2.

16. Description of Pass ill

Pass III reads a sequence of data statements that have been ordered
according to increasing action times, and it executes the operations
specified by these data statements. The principal operations are defining
instruments and playing notes. In addition, functions, variables, and
numbers may be computed and stored in the Pass III memory for
subsequent use in playing notes.
As mentioned in the introduction, most of the data in Pass III are
stored in a large linear array 1. Included are instrument definitions,
input-output blocks for unit generators, functions, note parameters.
The size of the various parts of I will vary greatly, depending both on
the specific computer being used and on the composition being played.
Consequently the structure of I is described in the IP data array, which
may be easily changed. Details are given in Section 17.

Over-all Operation
The over-all operation of Pass III is diagrammed in Fig. 56. The
program is started by reading a few constants from IP, including the
sampling rate IP(3) and the scale factor for variables IP(12).
A section is started by resetting the "played to" time T(l) to zero,
since time is measured from the beginning of each section.
The main loop of Pass III consists simply of reading a data statement
into the P array. As in previous passes, the P array is used exclusively
for reading and processing data statements. The operation code always
appears in pel) and the action time in P(2). Samples of the acoustic
output are generated until the" played to" time equals the action time.
Then the operation code is interpreted and executed. The next data
statement is then read and processed.

Instrument Definitions
If the operation code defines an instrument, the definition is entered
in the I array starting with the first empty location in the table for
instrument definitions. The location of the beginning of this instrument
definition is recorded in the location table for instrument definitions.
Different instruments are designated by being numbered.
154 CHAPTER THREE

Enter Enter Set

instrument Generate variable End End
note
parameters definition function or section piece
integer

Fig. 56. Block diagram of main loop-Pass III.

An instrument definition consists of a list of the various types of unit

generators used by the instrument, together with the inputs, outputs,
and functions for these units. Inputs and outputs can be note parameters
obtained for each note from a data statement, variables that are single
numbers stored more or less permanently in the I array, or input-output
blocks. These blocks are used for intercommunication between unit
generators. Instrument definitions continue unchanged from one
section to the next-unless they are redefined, in which case the latest
definition applies.
Note Playing
If the operation code specifies a note to be played, the data from the
P array are moved into the first unused block of locations in the note-
 MUSIC V MANUAL 155

parameter storage area. Unused note-parameter blocks have -I in

their first location; otherwise this location contains the instrument
number. To scale amplitudes the P(2) ---* Pen) parameters are multiplied
by IP(12) before storage in the note-parameter block (see paragraph on
scale factors, p. 157). The number of note-parameter blocks determines
the maximum number of voices that may be played simultaneously.
The termination time of the note is entered into the first unused
location of the TI array, and the I array subscript of the note param-
eters is entered in the corresponding location of the ITI array. Un-
used locations in TI are marked with the number 1000000.0. TI and
ITI are used to control the synthesis of samples of the acoustic
waveform.

Play to Action Time

The most intricate part of Pass III consists in generating acoustic
samples until the "played to" time T(I) equals the action time P(2) of
the current data statement. In the process, any notes terminating before
P(2) are turned off at their proper termination times. Several steps are
involved. This part of the program is diagrammed in Fig. 57.
The action time P(2) is put in the current "play-to" objective T(2).
The TI array that contains the terminating times of the instruments
currently playing is searched for the minimum termination time
TMIN. If TMIN < T(2), acoustic samples are generated until T(1) =
TMIN, TMIN is removed from TI, and the whole process is repeated.
If TMIN > T(2), samples are generated until T(1) = T(2) and control
is returned to the operation-code interpreter (F0RTRAN statement
200). If TI is or becomes empty, a rest is generated until T(l) = T(2).
The algorithm just described starts at F0RTRAN statement 244 as
indicated in Fig. 58.
The playing routines start at statement 260. The number of samples
to be generated ISAM is computed as the product of the sampling rate
1(4) times the time to be currently generated T(3) - T(1). T(3) is the
current objective for T(I). Sample generation proceeds by blocks. The
length of the block is the minimum of (I) the length of a unit-generator
input-output block, or (2) the number of samples remaining in ISAM.
For each block of samples the program scans all note-parameter blocks
(statement 268). For each voice that is turned on (first note parameter I:
-I) the program scans the instrument definition specified by the first
note parameter (statement 271). Each unit generator specified in the
instrument definition is called in the order in which it occurs in the
definition. Either SAMGEN or F0RSAM is called, depending on
156 CHAPTER THREE

ISAM =[T(3)-TtI)]*
[sampling rate]
T(I)=T(3)

271
Go through Go through
note unit
parameter generator
blocks in list for
Remove TMIN I array the given
from TI. Instrument
remove note
parameters
in I

(0 )

( b)

Fig. 57. Expansion of complicated parts of Pass III: (a) expansion of

"play to action time," block 244; (b) expansion of "play samples to T(3),"
block 260.

whether the unit-generator number is an integer less than or greater

than 100.5.
After all unit generators in all instruments have produced a block of
 MUSIC V MANUAL 157

samples, the block is outputted with SAM0UT (statement 265).

Another block is generated until ISAM is reduced to zero.
Function Generation
Some unit generators, in particular oscillators, use stored functions.
These are computed and stored by F0RTRAN subroutines GENI,
GEN2, . .. which may be supplied by the user. Existing GENn func-
tions for computing straight-line functions and sums of sinusoids are
described in Section 25. Upon reading a data statement that requests
function generation, Pass III calls upon the requested function. Space
to hold the functions is provided in the I array.
Scale Factors
Because F0RTRAN stores only integers in fixed-point arrays such
as I, variables that are inputs to unit generators are multiplied by
IP(12), which is set equal to 2n. This is equivalent to putting the decimal
point n places from the right end of the memory word. For a machine
with a 36-bit word, n is typically 18. Likewise functions are multiplied
by IP(15) which is typically set to 235 - 1 in a 36-bit machine.
Variable and Number Storage
The operation codes SV3, SI3, and SIA cause numbers and variables
to be stored in the I array. Variable number 1 is stored at 1(101),
number 2 at 1(102), etc. The appropriate P field is multiplied by the
scale factor IP(12) before storing. Thus I(m) equals IP(12) * Pen).
Integers are stored starting with integer 1 at 1(1), integer 2 at 1(2), etc.
In general, these numbers are used to control the program. The follow-
ing locations in I have special uses
1(4) Sampling rate
1(7) Master random number
1(8) Mono-stereo control. 1(8) = 0 for monophonic output,
1(8) = 1 for stereophonic
These may be changed as desired. Otherwise 1(1) through 1(20) are
reserved for program control and should not be changed. The SI3 and
SIA operations do not use a scale factor.
Multiple-Use Instruments and Unit Generators
The structure of Pass III has been designed so that the same block of
code embodying a unit generator is used in all instruments. Furthermore
the same instrument can simultaneously (in the sense of time of the
acoustic wave) produce many voices. This requires that no data specific
to a given instrument or voice can be stored in the unit-generator code.
158 CHAPTER THREE

Note-parameter blocks in the I array are kept intact for the duration of
a note. Hence certain quantities that must be continuous throughout
the note, particularly SUM in the oscillator, should be kept in the
note-parameter block.
Input-output blocks for unit generators must not be incorrectly
overwritten inside an instrument. The same block may be used as
input and output to a given unit generator, since the input is read before
the output is written. However, a block cannot be used simultaneous1y
for two different purposes, for example, as two inputs to a unit generator.
That is, it should be kept in mind that an input-output block may
contain only one set of values at a time (see Fig. 58).

(a)
( b)

Fig. 58. Examples of (a) an incorrect and (b) a correct input-output block.

17. I and IP Data Arrays in Pass III

Most of the data in Pass III are kept in a large one-dimensional
array I(n). Included are instrument definitions, note parameters,
functions, unit-generator input-output blocks, and a few other miscel-
laneous data. Except for a few fixed locations which will be listed below,
the data arrangement is flexible and is determined by parameters
compiled into the IP(n) parameter table. IP contains the main Pass III
constants which may change from time to time or from one computer
to another-constants such as the number and size of the functions,
scale factors for variables and functions, the sample value equal to zero
pressure in the acoustic output wave, etc.
The I array is usually structured as shown in the diagram below.
Values stored in IP give the subscripts in I(n) at which various quantities
are stored.
 MUSIC V MANUAL 159

Integers, variables, and special parameters

Variable n is located at I(n + 100)
}
Integer n is located at I(n)

IP(2) ~ Functions (produced by GEN subroutines)

} Function n begins at
I(lP(2) + (n - I) * IP(6))

IP(13) ~ Input-output blocks for unit generators

Block n begins at
}
I(IP(13) + (n - 1) * IP(l4))

IP(7) - - .
} Note parameters

IP(5) --.
Location table for instrument definitions
The definition of instrument n begins at
}
I(I(lP(5) + n))

IP(4) - - - .
} Instrument definition table

For example if IP(2) = 1000, functions will start at 1(1000).

Certain special parameters in I have fixed locations and a particular
meaning, as follows

I(1) Number of words on the current data statement in the P(n)

array
1(2) Subscript of first empty location in instrument definitions
1(3) Subscript of note parameters for the note currently being
played
1(4) Sampling rate
1(5) Number of samples to synthesize in the current group
1(6) Subscript of starting location in the instrument definition
for the unit generator currently being played
1(7) Master random number
1(8) Monophonic-stereophonic signal
1(8) = 0 for monophonic; 1(8) = 1 for stereophonic
160 CHAPTER THREE

Any location in the I array may be set by an SV3, SI3, or SIA

operation. In the set-variable operation the scale factor for variables is
used so that

I(n) = IP(l2)· P(m)

whereas for integers no scale factor is involved

I(n) = P(m)

The following constants are compiled into the IP(n) array. The array
is constructed by a BL0CK DATA subprogram and is stored in labeled
C0MM0N memory, labeled PARM.
IP(I) Number of operation codes in Pass III
IP(2) Beginning subscript of functions
IP(3) Standard (default) sampling rate
IP(4) Beginning subscript of instrument definitions
IP(5) Beginning subscript of location table for instrument
definitions
IP(6) Length of a function
IP(7) Beginning subscript of blocks of note-parameter storage
IP(8) Length of a block of note-parameter storage
IP(9) Number of blocks of note parameters (equals the maxi-
mum number of voices that can play simultaneously)
IP(lO) Subscript of unit-generator input-output block which is
reserved for storage of samples of the acoustic output
waveform. SAM0UT puts out samples from this block
IP(lI) Sound zero. This is integer with decimal point at right end
of the word
IP(l2) Scale factor for unit-generator variables (input-outputs,
etc.)
IP(l3) Subscript of beginning of unit-generator input-output
blocks
IP(I4) Length of a unit-generator input-output block
IP(l5) Scale factor for functions

18. Note Parameters

The word count and parameters PI through Pn are read by Pass III
from a data statement on the input file and are initially put into 1(1)
and into the P array. If PI = 1 == N0T, the parameters must be moved
 MUSIC V MANUAL 161

to a vacant block of note-parameter storage because other data state-

ments will be read into the P array before the N0T is completed. Note-
parameter blocks start at I(n) (n = IP(7)), each block is IP(8) locations
long, and IP(9) blocks are available.
A block contains the following arrangement of the information
I(n) = P3 (the instrument number)
I(n + 1) = P2 * IP(l2)
I(n + 2) = P3 * IP(l2)
I(n +m+ 1) = Pm * IP(12)
All subsequent locations to end of block are filled out with zeros. All
locations are in fixed-point format. All locations except the first are
scaled by the IP(12) scale factor. The first location I(n) contains the
instrument number, unscaled. If a block is empty, I(n) contains -1.
When a unit generator is called to calculate part of a note, 1(3) = n
is set to the first location of the note-parameter block for that note.
Consequently note parameter Pk may be found at I(n + k - 1).

19. Instrument Definition

An instrument in Pass III is defined by a sequence of data statements,
which are read from the input medium. The description is stored in the
I(n) array in the instrument definition table.
The format of the input data statements in Pass III is in the following
table.
Word
Record # Count P(l) P(2) P(3)

1 3 2 Action time Inst No

2 n 2 Action time Unit type D 1 .•• D n - 3
3 n 2 Action time Unit type D 1 •..

last 2 2 Action time

The description is terminated by a two-word statement. The quantities

Di specify the various inputs and outputs to the unit generators.
If Di < -100, then JDiJ - 100 is a function number
If -100 :::; Di < 0, then JDiJ is the number of a unit-generator
input-output block
If 1 :::; Di :::; 100, then Di is a note-parameter number
If 100 < Db then Di - 100 is a variable number
162 CHAPTER THREE

The mnemonic form of instrument definition as written on the score

and read by READ 1 has already been described in Section 4. Examples
are given in Section 5.
The instrument definition is stored starting in I(n) where n =
1(IP(5) + Inst number). Instrument definitions are stored in successive
locations in I(n) according to their action times. The first unused
location in the instrument definition table is kept in 1(2). An instrument
with a given number may be redefined at any action time. The new
definition will be used subsequently. However, no " garbage collection"
is done and the old definition will continue to occupy space in I(n).
The format of the description in I(n) is as follows:
I(n) Type of first unit generator in instrument
I(n + 1) Pointer to second unit-generator description = m

S1 )
~~.
Subscripts and parameters pertaining to first unit
generator
Sn
I(m) Type of second unit generator
Pointer to third unit generator
/.
I(r) 0 Terminates description of instrument
The S/s that specify inputs, outputs, and functions for the unit
generators have the following meaning:
If SI < 0, then /SI/ is the subscript in I which specifies the beginning
of a function or of a unit generator input-output block.
If 1 ::::; SI ::::; 262,144, then SI is the number of a note-card parameter.
If 262, 144 < Sh then SI - 262,144 is the subscript in I of a variable.
The number of the variable is SI - 262,144 - 100.

20. F0RSAM
F0RSAM is a subroutine that contains unit generators written in
F0RTRAN. These may be used either sepa,rately or together with
SAMGEN which contains unit generators written in basic machine
language.
F0RSAM is called in Pass III by the statement
CALL F0RSAM
The call causes F0RSAM to compute NSAM (= 1(5» samples of the
 MUSIC V MANUAL 163

output of unit generator, type J (J is given in the instrument descrip-

tion). Unit-generator types in F0RSAM are numbered 101 and up in
order to differentiate them from SAMGEN unit generators, which are
numbered 1 through 100.
The block diagram of the program is shown in Fig. 59.

Common
initialization
af addresses
and parameters

Fig. 59. Block diagram of

F0RSAM program.

Computation can be made in either fixed or floating-point arithmetic.

Usually the scale-factor variables, IP(12) and IP(15), will be necessary
to scale the results.
A listing of a small F0RSAM program with only one unit generator
is shown below. The initializing routines in the program accommodate
other unit generators which can be added to the program simply by
extending the G0 T0 at 205 and writing the unit-generator code.
The dimension statement includes three arrays from Pass III, I, P,
and IP, and two other arrays, Land M, which are used to address the
unit-generator inputs and outputs. Land M are computed by the
initialization procedure.
Specifically, the jth input or output will be found at I(m) where
m = L(j). M indicates whether an input or output is a single number
(note parameter or variable) or a block of numbers (function or 1-0
block). If M(j) = 0, the jth input is a single number; if M(j) = 1, the
jth input is a block. For blocks, L(j) gives the subscript of the first
number in the block. Inputs and outputs are sequentially numbered.
Thus in the data statement
0SC P5 P6 B2 Fl S ;
P5 is the first, P6 the second, B2 the third, Fl the fourth, and S the
fifth. For more convenient referencing, an equivalence is set up so that
L(i) == Li and M(i) == Mi.
164 CHAPTER THREE

SUBROUTINLFOHSAM
DIMENSIONI(15000),P(100),IP(20),L(8),M(8)
COMMONI,P/PARM/IP
LQUIVALENCE(Ml,M(I»,(M2,M(2»,(M3,M(3»,(M4,M(4»,(M5,M(5»,(M6,M
1(0», 0i07,~H7», (Ma,tJI(B», (L1,L(!», (L2,L(2», (L3,L(3», (L4,L(4», (
2L5,L(5) ), (L6,Ub) ) , (L 7, U7) ) , (LB,·L (8) )
C COMMON INITIALIZATION OF GENERATOHS
N1=I(6) +2
N2=I (N1-1>-1
U0204J1=IJI , N2
J£=J1-~J1 + 1
IF(I(Jl»200,201,201
200 L(J2)=-!(Jl)
M(J2)=1
GOT02U4
201 M(J2)=0
IF(I(J1)-262144)202,~02,203
202 L(J2)=I(Jl)+1(3)-1
GOT0204
203 L(J2)=I(Jl)-262144
204 CONTINUE
NSA",1=I (5)
N3=I (lH-2)
WGEN= N3 -100
20~ 60TO(101,30G,300),N0EN
C UNIT GENERATOR 101- INTERPOLATING OSCILLATOR
101 SFU=IP(12)
SFF=IP(15)
SFUI=l./SFU
SFFI=l./SFF
SFUFI=SFU/SFF
SUM=FLOAT(I(L5»*SFUI
IF(Ml)21U,210,211
210 AMP=FLOAT(I(Ll»*SFUI
211 IF(M2)212,212,213
212 FREQ=FLOAT(I(L2»*SFUI
213 XNFUN=IP(6)-1
D0223J3=1,NSAM
J4=INT<SUM) +L4
FRAC=SUM-AINT(SUM)
216 Fl=FLOAT(I(J4»
F2=FLOAT(I(J4+1»
217 F3=Fl+(F2-Fl)*FRAC
IF(M2)21b,218,219
218 SUM=SUM+FREQ
60T0220
219 J4=L2+J3-1
SUM=SUM+FLOAT(I(J4»*SFUI
220 IF(SUM-XNFUN)215,214,214
214 SUM=SUM-XNFUN
215 J5=L3+J3-1
IF(M1)221,221,222
221 I(J5)=IFIX(AMP*F3*SFUFI)
60T0223
222 J6=L1+J3-1
I(J5)=IFIX(FLOAT(I(J6»*F3*SFFI)
223 CONTINUE
I(L~)=IFIX(SUM*SFU)
300 RETURN
E.ND

The number of samples to be generated is put in NSAM. Most of the

unit generators will operate with a loop such as D0 223 J3 = 1,
NSAM.
In the computations performed by the unit generator, it is necessary
 MUSIC V MANUAL 165

to test to see whether an input is a single number or an 1-0 block. If

Mj = 0, the jth input need only be obtained once from I(Lj). If the jth
input is an 1-0 block (M j = 1), then each value is obtained with the
help of the main D0 index J3. For example, the third input is located at
1(15) where

J5 = J3 + L3 - 1
The particular unit generator is an oscillator that interpolates between
adjacent values of the function (see Section 6 for discussion of why
interpolation is useful). Computations are carried out in floating-point
arithmetic. Since the input data are fixed-point numbers, they must be
floated and scaled by appropriate constants. Scale factors for 1-0
blocks and for functions are given in IP(12) and IP(15), respectively.
The necessary scaling constants are computed at 101.

21. SAMGEN
SAMGEN is one of the few basic machine language programs in
Music V. Consequently it must be written specifically for the particular
machine on which it is to be used. The Bell Laboratories program is
written in GMAP for a General Electric 635 computer. A few com-
ments about the program may be of use in designing programs for
other machines.
SAMGEN includes the unit generators of type numbers less than
100. The computation of the actual acoustic samples, which is the
preponderance of the computation in Music V, is done by SAMGEN.
The general form of SAMGEN is shown in Fig. 60.
SAMGEN is written in such a way that one procedure can be used to
set the parameters in all of its unit generators. This procedure accesses
the I array in common storage during Pass III in order to find out

1(3) the subscript in the I array of the note parameters for the
note being played
] (5) the number of samples to generate and
](6) the subscript in the I array for the instrument definition table
of the unit generator being played.

The procedure then reads through the instrument description for the
unit generator being played. (See instrument description, Section 19.)
For each unit generator, the procedure expects a certain number of
inputs CSt's) in a certain order, e.g., if unit type = 2 (oscillator), then
166 CHAPTER THREE

I ni tialize unit
gene rotor
being played

Fig. 60. Block diagram of SAMGEN program.

Sl = amplitude, S2 = frequency, S3 = output, S4 = function, and

S5 = sum. It then sets addresses in the specified unit generator accord-
ing to the following conventions:

If Si < 0, then ISil is the subscript in the I array of the beginning

of a function or of unit-generator 1-0 block.
If 0 < Si < 262,144, then Si is the number of a note parameter.
Note parameter Px is located at 1(1(3) + x-I). (See Section
18 for more information on note-parameter storage.)
If Si > 262,144, then Si - 262,144 is the subscript in the I array
ofa variable: variable x is located at I(x + 100).

After the addresses are initialized, SAMGEN transfers control to the

specified unit generator, which generates the number of samples
specified in 1(5).
The calling sequence is
CALL SAMGEN
Almost all information is supplied by the I array which is located in
unlabeled common storage according to the statement
C0MM0N I

SAMGEN uses no subroutines.

 MUSIC V MANUAL 167

22. SAM0UT
SAM0UT is another GMAP subroutine called by Pass III which
(1) scales samples which are ready to be output, and (2) calls FR0UT
to output these samples onto magnetic tape. Samples (S1) are scaled
according to
SI = SI/218 + 2048
The calling sequence is
CALL SAM0UT (IARRAY, N)
where IARRAY = address of first sample to be output, and N is the
number of samples to be output.
Other routines used by SAM0UT are
FR0UT4
No common storage is used.

23. SAM0UT for Debugging

This version of SAM0UT (cf. Section 22) is provided for debugging
purposes only. It is called by Pass III with the call
CALL SAM0UT (IARRAY, N)
in order to print out N samples starting from location IARRAY. It
must perform the same de scaling operations as the normal SAM0UT,
i.e.,
samplel = (sample1/2 18) + 2048
This version of SAM0UT is written in F0RTRAN and will print the
sample values in any convenient format. It is recommended that in
using this version of SAM0UT one should be careful of excessive
output since it is easy to ask for a very large number of acoustic
samples.

24. Acoustic-Sample Output Program: FR0UT

The subroutine package FR0UT is called by both Pass III and
SAM0UT in order to write the actual acoustic samples generated by
Music Vanta magnetic tape. FR0UT is coded in assembly language
rather than F0RTRAN (1) for efficiency and (2) because it must write
168 CHAPTER THREE

special physical records onto tape in a form suitable for digital-to-

analog conversion. This is usually not possible in a compiler language
such as F0RTRAN.
The exact form of FR0UT will depend on the particular machine
configuration of a computer installation. It is therefore necessary that
this program be written by an experienced programmer at any compllter
installation that desires to run Music V.
There follows a general description of the FR0UT programs written
at Bell Telephone Laboratories for use with the General Electric
GE645 computer. It should act only as a model for such a program
written for another machine.
Basically FR0UT simply takes sample values that are produced by
Music V, packs several samples into one computer word, and writes
them onto magnetic tape in a form suitable for digital-to-analog
conversion.
At BTL, the digital-to-analog converters operate with 12-bit samples.
Since the GE645 computer is a 36-bit word-length machine, FR0UT
packs the acoustic samples three per word.
One packed computer word is of the form
I 36 bits I
aaaaaaaaaaaab bbbbbbbbbb bcccccccccccc
L sample 1--1 L sample 2-' L sample 3.J
Since the maximum integer value that can be represented in 12 bits is
4095 10 , FR0UT screens the sample values it receives from Music V to
be sure that it falls in the range 0 to 4095. Should any samples to be
written by FR0UT be outside this range, they are clipped to 0 and 4095.
Pass III first calls FR0UTO during its initializing sequence with the
call
CALL FR0UTO (66,167)
where 66 is a file code (i.e., a logical file name of the tape file onto
which packed acoustic samples are to be written), and 167 is the record
length in 36-bit words to be written onto this tape (samples per tape
record = 3 x words per tape record).
Whenever Music V has produced some samples that are ready to be
output, subroutine SAM0UT is called by Pass III, which in turn calls
FR0UT with the call
CALL FR0UT4 (lA, N)
which writes N samples onto tape starting from the location IA.
 MUSIC V MANUAL 169

At the end of the composition, Pass ITT calls FR0UT with the call
CALL FR0UT3
FR0UT3 completes the output buffer, if it was only partially filled,
with zero-voltage samples, empties this last buffer onto tape, and
writes an end-of-file mark.
Packing of samples can be accomplished by machine-language
shifting instructions and buffering. Acoustic sample tapes typically are
unlabeled and unblocked, and use fixed-length records.
FR0UT3 prints a statement giving the number of samples out of
range in the file which has just been terminated.

25. GEN-Pass III Function-Generating Subroutines

GENI
GENI is a F0RTRAN subroutine to generate functions composed of
segments of straight lines. The calling sequence is
CALL GENI
Data are supplied by the Pen), I(n), and IP(n) arrays. The jth function
Fj(i) is generated according to the form shown in the diagram below.

'. TM

~ I I
I I
I I
I I

Linear interpolation is used to generate the function between M

points which are specified by the user. Thus between any two abscissa
points N m and Nm+ 1 the function points are computed according to the
relation

The number of corners M is arbitrary and is determined by the word

count 1(1). M = (1(1) - 4)/2.
170 CHAPTER THREE

In general the user will set NI = 0 and NM = IP(6) - 1, so that the

number of points in the function equals IP(6).
The parameters of the function are arranged as follows:
P(1) P(2) P(3) P(4) P(5) P(6) P(7) P(8)

3 Action Function T I N1
time No (j)

The function is stored starting in I(n) where n = IP(2) + (j - 1) *

IP(6) and is scaled by IP(15) so that, for example, len) = Tl * IP(15).

GEN2
GEN2 is a F0RTRAN subroutine to generate a function composed
of sums of sinusoids. The calling sequence is
CALL GEN2
Data are supplied by the Pen), I(n), and IP(n) arrays.
The jth function Fli) is generated according to the relation

F;(i) ~ (amp normaIiZer)t~, Ak sin ;~\

+ k~
M
Bk
2k"}
cos P-=- \ i = O... P - 1

P (= IP(6» is the number of samples in a function.

The parameters for the function are arranged as follows:
P(1) P(2) P(3) P(4) P(5) P( -) P( -)

3 Action 2 Function Al ±N
time No (j)

The number of sine terms is INI. If N is positive, amp normalizer

is computed so max IFli) I = .99999. If N is negative, amp normalizer
= .99999. The number of cosine terms M is computed from N and the
word count 1(1). M = I{l) - N - 5.
The number of samples in the function is IP(6).
The function is stored s~arting in I(n), and is scaled by IP(5)
= IP(15) * Fj(O), etc.
len)
where n = IP(2) + (j - 1) * IP(6).
Thefirst and last samples of the function are equal, Fj(O) = Fj(P - I),
thus the period in samples is P - 1.
 MUSIC V MANUAL 171

GEN3
General description:
GEN3 is a F0RTRAN subroutine which generates a stored function
according to a list of integers of arbitrary length. These integers specify
the relative amplitude at equally spaced points along a continuous
periodic function. The first and last points are considered to be the
same when the function is used periodically (e.g., by an oscillator).
Calling sequence:
CALL GEN3
Other routines used by GEN3 :
none
Data statement:
GEN, action time, 3, stored function number, PI, P 2 , • . . , Pnj
Examples:
The following P/s will generate the functions shown below.

(1) 0, 1,-1
o~----------------------~-------
will generate

(2) 0, 8, 10, 8, 0,
O~--------------~r-------------4
-8, -10,0:

(3) -1000, 0: o~------------------------~~~-

172 CHAPTER THREE

26. Data Input for Pass ill-DATA

Subroutine DATA is called by Pass III with the call
CALL DATA
This causes one data statement to be read from file 11 into the P array
in C0MM0N storage according to
READ (11) K, (P(J), J = 1, K)
I(I) is set equal to K (word count).

Annotated References by Subject

Music IV Program
M. V. Mathews, "The Digital Computer as a Musical Instrument," Science,
142, 553-557 (November 1963). A semitechnical description of Music IV
with some discussion of applications. This is a good introductory article.
M. V. Mathews, "An AcoustIc Compiler for Music and Psychological Stimuli,"
Bell Sys. Tech. J. 40, 677-694 (May 1961). A technical description of an
early version of a sound generating program. This is the first complete
published description.
J. R. Pierce, M. V. Mathews, and J. C. Risset, "Further Experiments on the Use of
the Computer in Connection with Music," Gravesaner Blatter, No. 27/28,
92-97 (November 1965). A semitechnical description emphasizing applica-
tions of Music IV. This is a good follow-up for the paper in Science.
J. C. Tenney, "Sound Generation by Means of a Digital Computer," J. Music
Theory, 7, 25-70 (1963). A discussion of Music IV as seen by a composer
using the program. The article contains many details and is a good introduc-
tion for a musician.
Appendix A Psychoacoustics
and Music
J. R. Pierce
and M. V. Mathews

Although the technology of electronic and computer sound generation

has given us new tools of almost unlimited power for making new
sounds, it has also created a new problem-the need to understand the
psychoacoustics of musical perception. Sounds produced by conven-
tional instruments are so well known that composers can proceed with
the intuitions they have developed from long experience. However, no
such intuitions exist for new sounds. Instead, the composer must
understand the relation between the physical sound wave and how it is
perceived by a hearer. Psychoacoustics addresses this question and
hence has become an essential knowledge for the modern composer.
With some exceptions (Helmholtz, 1863; Plomp, 1966) original
scientific work in psychoacoustics has not been directed chiefly at
musical problems. Thus we must draw on a variety of sources in
seeking to understand musical phenomena, and we may often wish that
investigators had had music in mind.

Loudness
The perceived loudness of a sound depends on many factors in
addition to its intensity. For example, in order for a pure tone or
sinusoid at 100 Hz to be heard, its sound intensity must be 1000 times
greater than that of a pure tone at 3000 Hz. For most of the musical

173
174 APPENDIX A

range the perceived loudness increases as the 0.6 power of the sound
pressure (Stevens, 1961). The perceived loudness increases more slowly
with sound pressure for 3000-Hz tones than it does for very low fre-
quencies, say, 100 Hz; and in the 'uncomfortably loud range, tones of
equal power are about equally loud. This means that as we turn the
volume control up or down, the balance of loudness among frequency
components changes slightly.

Masking and Threshold Shifts

A tone or a noise masks or renders us incapable of hearing a less
powerful tone. A tone has a strong masking effect for tones of higher
frequency and a weaker masking effect for tones of lower frequency.
The frequency range of masking is greater for loud tones than for soft
tones. Thus we would expect that in a musical composition some
sounds might be masked and unheard when the volume is set high,
whereas they would be unmasked and heard when the volume is low.
Masking can be considered as a raising of the level at which tones
become audible. Some rise in the threshold persists for i sec or longer
after a loud tone (Licklider, 1951), but the aftereffect of a loud tone on
hearing is much less than that of a bright light on seeing.

Limens or Just Noticeable Differences

Limens or jnd's of loudness and frequency have been carefully
measured. They are surprisingly small. However, there is evidence that
the limens are much smaller than the frequency or loudness differences
that can be detected in complicated listening tasks, which are more akin
to music (Plomp, 1966, p. 19). Very small differences in frequency (less
than a half tone) and loudness can be detected in successive tones that
are not too short.

Pitch

The pitch of a complex tone is often thought of as that of its lowest

partial. However, experiments made with repetitions of various patterns
of pulses (Flanagan and Guttman, 1960) and with complex tones in
which the upper partials are harmonics of a frequency higher than the
fundamental (Plomp, 1967) show that, although the fundamental
dominates at higher frequencies, the repetition rate of the tone or of its
higher partials dominates at lower frequencies. The pitch of a tone may
 PSYCHOACOUSTICS AND MUSIC 175

be highly uncertain by one or more octaves; thus Shepard produced a

circle of 12 tones, which when cyclically repeated give the impression of
always rising in pitch, with no break (Shepard, 1964). Tones with
inharmonic partials, including gongs, bells, and tones specially syn-
thesized by computers (Mathews, 1963; Pierce, 1966) may produce a
sensation of pitch; a tune can be played on them. But the pitch may not
be the first partial; for example, the hum tone of a bell is not the pitch
to which the bell is tuned.

Quality or Timbre of Steady Tones

The sound quality or timbre of steady tones depends on the partials.
Although partials up to the sixth (and sometimes higher) can be heard
individually by careful listening, we tend rather to hear an over-all
effect of the partials through the timbre of the tone. A pure tone or
sinusoid is thin. A combination of octave partials is bright. A tone with
a large number of harmonic partials is harsh or buzzy (Pierce, 1966).
In general, the timbre appears to be dissonant or unpleasant if two strong
partials fall within a critical bandwidth, which is about 100 Hz below
600 Hz and about a fifth of an octave above 600 Hz (Plomp, 1966).
The timbre of a sound is strongly affected by resonances in the vocal
tract or in musical instruments. These resonances strengthen the
partials near the resonant frequencies. Three important formants or
ranges of strengthened frequency are produced by the vocal tract; they
give the qualities to vowel sounds which are identifiable independent of
pitch.

Transient Phenomena
Textbooks give harmonic analyses of the sounds of various musical
instruments, but if we synthesize a steady tone according to such a
formula it sounds little like the actual instrument. Steady synthesized
vowels do not sound like speech if their duration is long.
Temporal changes such as attack, decay, vibrato, and tremolo,
whether regular or irregular, have a strong effect on sound quality. A
rapid attack followed by a gradual decay gives a plucked quality to any
waveform. Also, the rate at which various partials rise with time and
the difference in the relative intensity of partials with loudness are
essential to the quality of the sound (Risset, 1965). Indeed it is at least
in part the difference in relative intensity of partials that enables us to
tell a loud passage from a soft passage regardless of the setting of the
 176 APPENDIX A

volume control. This clue is lost in electronic music if the tones employed
have a constant relative strength of partials, independent of volume.
The "warmth" of the piano tone has been shown to be due to the
fact that the upper partials are not quite harmonically related to the
fundamental (Fletcher et at., 1962).

Consonance
Observers with normal hearing but without musical training find
pairs of pure tones consonant if the frequencies are separated by more
than the critical bandwidth (Plomp, 1966), or if the frequencies coincide
or are within a few hertz of one another (in this case beats are heard).
Pairs of tones are most dissonant when they are about a quarter of a
critical bandwidth apart. For frequencies above 600 Hz, this is about a
twentieth of an octave.
Excluding bells, gongs, and drums, the partials of musical instruments
are nearly harmonic. When this is so, for certain ratios of the frequencies
of fundamentals, the partials of two tones either coincide or are well
separated. These ratios of fundamentals are 2:1 (the octave), 3:2 (the
fifth), 4:3 (the fourth), 5:4 (the major third), and 6:5 (the minor third).
Normal observers find pairs of tones with these ratios of fundamentals
to be more pleasant, and intervening ratios less pleasant (Plomp, 1966).
Musical consonance and dissonance depend on many factors in
addition to frequencies of partials. For example, unlike nonmusicians,
classically trained musicians describe pairs of pure tones with these
simple numerical ratios of frequency as consonant and intervening
ratios as dissonant. The only reasonable explanation is that trained
musicians are able to recognize familiar intervals and have learned to
think of these intervals only as consonant.
Plomp (1966) has pointed out that, in order for complex tones to
attain a given degree of consonance, low tones must be separated by a
larger fraction of an octave than high tones, and he has observed that
composers follow this principle.
If the partials of a tone are regularly arranged but not harmonic, the
ratios of frequencies of the fundamental (or first partial) that lead to
, consonance are not the conventional ones (Pierce, 1966).

Combination Tones
When we listen to a pure tone of frequency f1 and another tone of
somewhat higher frequency f2' we hear a combination tone of lower
 PSYCHOACOUSTICS AND MUSIC 177

frequency 2f1 - f2' even at low sound levels (Goldstein, 1967). At

much higher sound levels, around 100,000 times or more the power at
threshold, it is possible to hear faint frequencies 2f1, 2f2, f1 + f2'
f2 - fb etc. Combination tones are due to nonlinearities in the hearing
mechanism. They can contribute to dissonance and to beats.

Reverberation
Reverberation is important to musical quality; music recorded in
an organ loft sounds like a bad electronic organ. The reverberation for
speech should be as short as possible; for music about 2 sec is effective.
Music sounds dry in a hall designed for speech. Reverberation is not the
only effect in architectural acoustics. Our understanding of architectural
acoustics is far from satisfactory (Schroeder, 1966).

The Choir Effect

Many voices or many instruments do not sound like one voice or one
instrument. Some experiments by the writers show that a choir effect
cannot be attained by random tremolo or vibrato. It must be due to
irregular changes in over-all waveform, caused by beating or head
motions, or by. differences in attack.

Direction and Distance

We can experience a sidedness to sound by wearing headphones fed
from two microphones, but the sound seems to be inside our head.
We experience externalization of the sound-as coming from a par-
ticular direction-only when we allow head movements in a sound field.
Although we cannot detect the direction of the source of a sinusoidal
tone in a reverberant room, we can detect the direction by the onset of
such a tone, and we can detect the direction of clicks and other changing
sounds. The first arrival of the sound dominates later reverberant
arrivals in our sensing of the direction of the source; this is called the
precedence effect (Wallach, Newman, and Rosenzweig, 1949). We can
detect vertical angle of arrival, although no one is sure how this is done.
We can also sense the distance of a source in a reverberant room; this
sensation must depend on some comparison of the direct arrival and the
reverberant sound (Gardner, 1967).

Memory and Overlearning.

Most memory experiments are not done with musical sounds, but
many are relevant to music.
178 APPENDIX A

Miller (1956) found that subjects can remember a sequence of from

5 to 9 randomly chosen digits, letters, or words. On the other hand, a
good bridge player can remember every card that has been played in
an entire game. Our ability to deal with stimuli depends on their
familiarity or "meaning" to us. This familiarity comes about through
overlearning. Overlearning has been insufficiently investigated because,
although it is common in life, it is very difficult to achieve in the
laboratory.
The phonemes of a language are overlearned. A subject can readily
distinguish the phonemes of his own tongue, but not those of another.
He can distinguish dialects of his own language, but not those of a
foreign tongue. He can understand his native language in a noisy place
better than he can understand a foreign language even though he is
expert in it.
Conventional elements and structures in music are undoubtedly
overlearned. Much of our appreciation of harmony, much of our ability
to remember conventional tunes (Mozart, Haydn, and some other
musicians could remember compositions heard only once) must
depend on overlearning, just as our ability to use and remember
language does. Performance with unfamiliar material is much poorer.

Psychological Distance; Scaling

Some psychological stimuli have the same pattern of similarity for
all people. Color is one. The psychological distance between stimuli
such as colors can be obtained by computer analysis of data expressing
either the confusions that subjects make among pairs of stimuli or the
numbers that they assign to the pairs to express their judgments of
similarity. This kind of analysis is called multidimensional scaling. The
stimuli may appear in a psychological space of one dimension (loudness
does), two dimensions (color does) or three (vowels do) or more
dimensions. Psychological distance is dependent on, but not propor-
tional to, physical parameters. Thus red and violet light are of all colors
the farthest apart in wavelength, and yet they look more alike-they
are closer together psychologically-than the "intermediate" colors
orange and blue.
Unhappily, multidimensional scaling is just beginning to be applied
in the field of music (Levelt et al., 1966). Further results might be
enlightening. For instance, we is nearly you said backwards, and yet we
perceive no similarity between the sounds of the two words. Is the
retrograde of a phrase psychologically similar to the phrase, or is
 PSYCHOACOUSTICS AND MUSIC 179

retrograde (in the words of Tovey) for the eye only? Transpositions
certainly are psychologically close, but what about augmentations and
inversions? What about changes in rhythm? What about manipulations
of the tone row?

References
Flanagan, J. L., and N. Guttman, "On the Pitch of Periodic Pulses," J. Acoust.
Soc. Amer. 32, 1308 (October 1960).
Fletcher, H., E. D. Blackham, and R. Stratton, "Quality of Piano Tones,"
J. Acoust. Soc. Amer. 34, 749 (June 1962).
Gardner, M., "Comparison of Lateral Localization and Distance for Single- and
Multiple-Source Speech Signals," J. Acoust. Soc. Amer. 41, 1592 (June 1967),
Abstract.
Goldstein, J. L., "Auditory Nonlinearity," J. Acoust. Soc. Amer. 41, 676-689
(March 1967).
von Helmholtz, H. L. F., Die Lehre von der Tonempfindungen als physiologische
Grundlage fur die Theorie der Musik, 1863. On the Sensations of Tone as a
Physiological Basis for the Theory of Music (Dover, New York, 1954).
Levelt, W. J. M., J. P. van de Geer, and R. Plomp, "Triadic Comparisons of
Musical Intervals," Brit. J. Math. Statist. Psychol. 19 (Part 2), 163-179
(November 1966).
Licklider, J. C. R., "Basic Correlates of the Auditory Stimulus," in Handbook of
Experimental Psychology, S. S. Stevens, Ed. (John Wiley & Sons, New York,
N.Y., 1951).
Mathews, M. V., "The Digital Computer as a Musical Instrument," Science 142,
553 (November 1963).
Miller, G. A., "The Magical Number Seven, Plus or Minus Two," Psycho!. Rev.,
63, 81 (1956).
Pierce, J. R., "Attaining Consonance in Arbitrary Scales," J. Acoust. Soc. Amer.
40, 249 (July 1966).
Pierce, J. R., and E. E. David, Man's World of Sound (Doubleday, Garden City,
N.Y., 1958).
Plomp, R., Experiments on Tone Perception (Institute for Perception RVO-TNO,
Soesterberg, The Netherlands, 1966).
Plomp, R., "Pitch of Complex Tones," J. Acoust. Soc. Amer. 41, 1526-1533
(June 1967).
Risset, J. C., "Computer Study of Trumpet Tones," J. Acoust. Soc. Amer. 38,
912 (November 1965), Abstract.
Schroeder, M. R., "Architectural Acoustics," Science 151, 1355 (March 1966).
Shepard, R. N., "Circularity in Judgments of Relative Pitch," J. Acoust. Soc.
Amer. 36, 2346 (December 1964).
Stevens, S. S., "Procedure for Calculating Loudness: Mark VI," J. Acoust. Soc.
Amer.33, 1577-1585 (1961).
Wallach, H., E. B. Newman, and M. R. Rosenzweig, "The Precedence Effect in
Sound Localization," Amer. J. Psychol. 52, 315-336 (1949).
Appendix B Mathematics

In the body of this text an effort has been made to minimize the number
and difficulty of mathematical expressions. In certain places some
computations characteristic of signal theory must be done. This
appendix lists the relations that are required by the text. No proofs are
given, and the conditions under which the relations are true are not
spelled out. They hold in a useful (and widely used way) for almost all
real signals. We apologize for the strong MIT and EE accent in the
mathematical language. If one has something to say, it is better to
speak with an accent than to remain silent.

Fourier Series
A "not too discontinuous" function f(x) with period T can be
represented almost everywhere by the series
a 277" 4rr
f(x) = "2o + a l cos y x + a2 cos y x + ...

. 277" b ' 4rr

+ b lSlnyX+ 2SlnyX+'"

where
2 fT 217-i
at = T J0 f(x) cos T dx

180
 MATHEMATICS 181

r .
and

bi = if2 0
27Ti
f(x) sm T dx

Fourier Transform
A "not too discontinuous" function f(x) for which the integral of
f2(X) exists may be transformed and inverse transformed according to
the relations
+OO
pew) =
f -00 p(t)e- Joot dt

1 f+oo P(w)eJoot dw
pet) = -27T _ 00
pew) is called the Fourier transform of pet); pew) is also called the
amplitude spectrum of pet).

Input-Output Relations for Time-Invariant Linear Systems

The output oCt) of a time-invariant linear system due to an input
i(t) may be written
+OO
oCt) =
f
_ 00 i(t - x)h(x) dx

where hex) is called the impulse response of the system. For realizable
systems, hex) = 0 for x < O. The transform of hex) is called the
transfer function H(w) of the linear system and is written
+OO
H(w) =
f _ 00 h(t)e- Jrot dt

The Fourier transform of the output O(w) and the Fourier transform
of the input I(w) are related by the simple equation
O(w) = H(w)I(w)

Convolution Theorem
The three time functions, x(t), yet), and z(t), have as their respective
Fourier transforms X(w), Yew), and Z(w). If z is the product of x and y
z(t) = x(t)· yet}
182 APPENDIX B

then
J+
1 _ 0000 X(a)Y(w - a) da
Z(w) = 27T

If Z is the product of X and Y

Z(w) = X(w). Yew)
then
+OO
z(t) =
J _ x(a)y(t - a) da
00

Definition of Unit Impulse

The unit impulse Set) can be considered the limit

Set) = lim 1 e -t2 /2u 2

u->O V27Ta
I t is zero everywhere except at t = 0, but its integral is unity
+OO

J_ 00 S(t) d t = 1

Spectrum of Cos wot

Although the integral of cos 2 wot does not exist, and hence cos wot
does not have a legitimate Fourier transform, the transform pew) can
be usefully defined as
pew) = 7T[S(W - wo) + sew + wo)]
Note in particular that the inverse transform

= cos wot

Autocorrelation Function and Power Spectrum

If pet) is an ergodic random function, then an autocorr~lation
function cp( T) may be defined by the relation

cp(T) = J~~ 2~ f~T p(t)p(t + T) dt

 MATHEMATICS 183

More generally
cp(1") = E<p(t)p(t + 1")
where E< ) is defined in some way that makes sense for the random
function pet). The power spectrum <I>(w) is the Fourier transform of
cp(1"). Thus

<I>(w) = -1 f+oo cp(1")e- jOlt dt

27T _ 00

cp(1") = s:: <I>(w)e jw • d1"

Note that the 2~ factor is in the transform rather than the inverse
transform.

Random Functions and Linear Systems

H(w) is the transfer function of a linear system having an input i(t)
and an output oCt). Let <l>i(W) and <l>o(w) be the power spectra of the
input and output, respectively; then
<l>o(w) = IH(w) I2<l>i(W)

Mean-Square Function
If pet) is a random function with autocorrelation function cp(1") and
power spectrum <1>(w), then
+OO
E<p(t)2) = cp(O) =
f _ <I>(w) dw
00
Index

AD2-AD4: adder units, 128 Cos mot, spectrum of, 182

AD2 generator, tutorial discussion, 47 D array
Algorithm, 135 Pass I, 81, 122
Amplitude function, 91 Pass II, 95, 146
Analog-to-digital converter, 26 DATA, operation of, 172
Attack, 53, 73, 175 Data statement
Autocorrelation function, 182 function of, 117
for Pass I, 121
Band-pass noise, 70, 129 Decay, 53, 73, 175
Beats, 87 Define instrument, INS, \23
Bn, 125 Digital data storage, 31
Buffer memory, 32 Digital-tape control for sound
recording, 32
CARD,139 Digital-to-analog converter, 27
Choir effect, 177 Direction perception, 177
C0M,121 Dissonance, 176
Combination tones, 176 Distance perception, 177
Composing subroutines, tutorial Duration function, 90
discussion, 78 Duration of note, P4, 46
Compositional functions, 86 Duty-factor function, 90
C,0N, function evaluator, 150
tutorial discussion, 91 End section, SEC, 123
Consonance, 176 ENV: envelope generator
Convolution theorem, 181 description, 130
C0NVT tutorial discussion, 73
convert subroutine, 152 ERR0R, operation of, 144
for ENV, 75 ' ERR~ codes, table of, 145
for glissando, 66 External data connection, 33
simple example, 63
tutorial discussion, 62 Filter, errors introduced by, 15
for 12-tone scale, 65 Filters, smoothing, 16

185
186 INDEX

FLT: filter, 76 IP array

Fn,125 Pass I, 81, 121
Foldover errors, 7, 15, 19 Pass II, 63, 148
F0RSAM Pass III, 119, 160
block diagram, 163 IP (12), 157
operation of, 162 IP (15), 157
F0RTRAN, 43 ISAM,157
Fourier series, 180 ITI array, 155
Fourier transform, 181 IVT,142
Frequency-limited function, 12
Frequency scale, logarithmic, 80 J in READl, 143
FR0UT, operation of, 167 Just noticeable differences, 174
FR0UTO, 168
FRfOVT3, 169 Lin READl, 143
FR0UT4,168 Limens, 174
Function evaluator, C0N,150 Linear systems with random inputs,
183
G array, Pass II, 66, 146, 148 L0P,142
GEN program for stored functions Loudness, 173
data statement, 123 LSG, line-segment generator, 100
use in Pass III, 157
use of, 50 Magnetic tape, digital, 31
GEN1, operation of, 169 Masking, 174
GEN2, operation of, 170 Mean-square function, 183
GEN3, operation of, 171 Memory and overlearning, 177
Glissando, simple instrument with, 58 Metronome function
Graphic score, 89 example, 87
operation of, 148
HARVEY, 121 MLT generator, tutorial discussion, 49
M0VL, operation of, 141
I array M0VR, operation of, 141
Pass II, 95, 146 Music I-Music V
Pass III, 119, 155, 159 general description, 34
I used by READ1, 143 references, 172
IBC,142 Music V
IBCD,139 comparison with Music IV, 115
ICAR,139 general operation of, 116
Impulse modulator, 12 overviewof,37
Input-output blocks, correct use of,
158 NBC, number of break characters, 142
Input-output routines for Pass I and N0T, play note, 123
Pass II, 139 Note concept, 34, 36
INS, define instrument, 123 Note-parameter storage blocks, 154,
Instrument with attack and decay, 53 155, 160
Instrument definition Note playing, operation of, 154
operation of in Pass III, 161 NPW, 143
table, 162 NSAM, 162, 164
Instrument that varies waveform with Numerical representation of functions
amplitude, 60 of time, 2
Instruments NUMU, 143
definition of, 124
interactions between, 98 OP codes
multiple use, 157 adding more, 142
1-0 blocks, 47 table of, 122
 INDEX 187

Orchestra, elementary example, 44 Psychoacoustics

0SC generator importance, 173
description, 127 references, 179
distortions in, 13 8 Psychological distance, 178
with interpolation, 138
with rounding, 138 Quantizing, 5
special discussion, 134 Quantizing errors
with truncation, 138 basic equations for, 24
tutorial discussion, 49 definition, 7
0UT generator mathematical analysis, 22
description, 126
tutorial discussion, 48 RAH: random and hold generator, 132
RAN: random function generator
description, 128
P array tutorial discussion, 68
Pass I, 81, 121 use for band-pass noise, 70
Pass II, 63 use for vibrato, 70
PARM,160 Random and hold generator: RAH,
Pass I 132
block diagram, 120 READ,U8
C0MM0N statement, 121 READO
description, 120 debug, 143
general operation, 116 operation of, 139
Pass I subroutines, 118 READ 1
Pass II block diagram of, 140
block diagram, 146 debug, 143
C0MM0N statement, 146 operation of, 139
description, 145 stereo-mono control, 142
G array, 148 READ2,143
general operation, 118 Realizable filters, errors in, 15
IP array, 148 Real-time synthesis, 34
report, 149 Report, Pass II, 149
Pass II subroutines, PLS, 94 Reverberation, 177
Pass III
block diagrams 154, 156 SAMGEN
description, 153 block diagram, 166
general operation, 119 operation of, 165
Pitch, 174 SAM0UT for debugging, 167
Pitch function, 91 operation of, 167
Pitch-quantizing example, 94 Sample and hold analysis, 21
Play note, N0T, 123 Samples of acoustic waveform, Si!, 46
Play-to-action time, 155 Sampling
PLF, execute subroutine in Pass I, 123 alternative analysis, 16
general operation, 144 basic equation, 14
score record, 82 definition, 4
use as composing subroutine, 78 error bounds, 18
use to multiply melodies, 83 errors, 15
use with graphic score, 89 mathematical analysis, 11
PLS, execute subroutine in Pass II, 124 Sampling interval, T, 11
operation of, 151 Sampling rate, R, 5, 11
tutorial example, 94 Sampling switch for digital-to-analog
Pn, 125 converters, 29
Power spectrum, 182 Scale factors, 157
Pressure function p (t), 2 Scaling, 178
188 INDEX

Score, elementary example, 44 T array, Pass II, 146

Score statements, description of, 117 T(1), 153, 155
SEC, end section, 123 T(2), 155
SET: set new function number, 134 T(3), 155
Set integer in all passes, SIA, 124 Tempo function, 148
Set integer in Pass III, SI3, 124 TER, terminate piece, 123
Set new function number: SET, 134 Threshold shifts, 174
Set variable in Pass I, SVl, 123 TI array, 155
Set variable in Pass II, SV2, 123 Timbre of steady tones, 175
Set variable in Pass III, SV3, 123 Time-invariant linear systems, 181
SIA, set integer in all passes, 124 Time scale, 34
SI3, set integer in Pass III, 124 TMIN,155
Signal-to-noise ratio from quantizing, Training orchestra, parameters, 105
7 Tremolo, 175
Smoothing filter Tutorial examples
design of, 29 problems, 106
use of, 4 references, 105
SNA8,142
S0RT, 146, 150 Unit generator
S0RTFL, 146, 150 description of, 34
Sound-processing fundamentals list of, 124
problems, 40 simple, 46
references, 39 Unit impulse, definition of, 182
Sound-synthesis programming, Variables
fundamentals of, 33 in Pass I, 81
Spectrum of sampling impulses, 13 in Pass II, 66
Starting time of note, P2, 46 in Pass III, 60
Steady-state time, 75 Vibrato
STER,142 simple instrument with, 55
Stereophonic output box: STR, 131 in tone quality, 175
Stored function, 34,49, 50, 135 using RAN, 70
STR: stereophonic output box, 131 Vn,125
Subroutine in Pass I, PLF, 123 Voices
Subroutine in Pass II, PLS, 124 combining, 46
SVl, set variable in Pass I, 123 definition, 36
SV2, set variable in Pass II, 123 Voltage function of time, 3
SV3, set variable in Pass III, 123, 157
Swell and diminuendo, simple Word size, changing, 142
instrument with, 58 WRITEl,143
Swells, instrument for, 99 WRITE2,148