Digital audio processing revisited

Juan P Bello
Digital audio processing
Microphones
• Sound is an energy disturbance that propagates through a
medium as a wave
• Commonly, the medium is air, thus the sound wave produces
variations of air pressure
• A microphone is a transducer (i.e. a device that converts
energy or information from one form to another).
• Specifically, the microphone converts air pressure into voltage
levels, thus generating an electrical signal analogous to the
mechanical one.
• The following expression notates the relationship between
voltage and pressure in a microphone, where the symbol µ
means “is proportional to”: v(t) µ p(t)
ADC
• The conversion of an analog (continuous) signal x(t) into a
discrete sequence of numbers x(n) is performed by an Analog-to-
digital Converter (ADC)
• The ADC samples the amplitude of the analog signal at regular
intervals in time, and encodes (quantizes) those values as binary
numbers.
• The regular time intervals are known as the sampling period (Ts)
and are determined by the ADC clock.
• This period defines the frequency at which the sampling will be
done, such that the sampling frequency (in Hertz) is:
1
fs =
Ts
• The accuracy of the quantization depends on the number of bits
used to encode each amplitude value from the analog signal.

!
ADC

ADC

• The outgoing sequence x(n) is a discrete-time signal with

quantized amplitude
• Each element of the sequence is referred to as a sample.

...,x[n " 1],x[n],x[n +1],...

Discrete signals
• An example discrete signal is a real sinusoid, which can be
described as:
x[n] = a cos("n + # )
• where a is the amplitude, ω the angular frequency, and φ the
initial phase. At sample number n, the phase is equal to φ + ωn.
• A sinusoid is an example of simple harmonic motion.
• Because each cycle is completed in a constant amount of time,
!
the motion of the wave is periodic, i.e. there is a T > 0 that
satisfies the equation:

f (n) = f (n + "),0 < " < #

• The number of cycles completed per second is the frequency of
the wave, and the inverse of the frequency is its period.
Discrete signals

• A sine and a cosine are differentiated only by a phase difference

of a quarter cycle (π/2)
The sampling theorem
• Sampling is the process of converting a continuous signal into a
discrete sequence
• Our intuition tells us that we will loose information in the process
• However this is not necessarily the case and the sampling
theorem simply formalizes this fact
• It states that “in order to be able to reconstruct a bandlimited
signal, the sampling frequency must be at least twice the highest
frequency of the signal being sampled” (Nyquist, 1928)

The Nyquist frequency

Aliasing
• What happens when fs < 2B
• There is another, lower-frequency, signal that share samples with
the original signal (an alias).

• Related to the wagon-wheel effect:

http://www.michaelbach.de/ot/mot_strob/index.html

LPF
Anti-aliasing
Hearing frequency range

Human hearing is widely

accepted to lie in the
20-20kHz range

Thus main reason for

standard sampling
frequencies to be of
44.1kHz and 48kHz

In digital synthesis we
then have to be careful
not to exceed the
Nyquist frequency
Loudness

• dB = 10 * log10(level/reference level) - Levels of intensity or power

• Reference level = 0dB = 10-12 watts per square meter (threshold of
hearing)
Dynamic Range
• Threshold of hearing is ~0dB and threshold of pain is ~125dB
• Dynamic range of a system: difference between the loudest and
softest sound that a system can produce (measured in dB)
• On a linearly encoded PCM streams it is roughly: # of bits * 6

Dynamic Range
Quantization noise
• Is the distortion produced by the rounding-up of real signal amplitude
values during the ADC process to the values “allowed” by the bit-
resolution of each sample.
• The difference in level between the intended signal and the noise arising
from quantization is the signal-to-quantization-noise ratio (SQNR)
• This depends on the quantization accuracy (# of bits) and the signal itself.

• Example: a sound with progressively worsening quantization noise:

Low-level quantization noise
• Sounds just above silence are degraded most severely by the
quantization noise, because all of the variation is captured by the
least significant bit.

• This is known as low-level QN, i.e. a square wave produced by 1-

bit variations triggered when the signal has a very low amplitude.
• This noise can be critical as square waves are rich in odd
harmonics, that can even extend beyond the Nyquist frequency
producing aliasing.

• Solutions to this problem include:

1. Increasing the bit resolution (the level of noise is “inversely
proportional” to the number of bits per sample)
2. Adding dither, i.e. low-energy analog noise added prior to the AD
conversion, hence randomizing the quantization noise. Low-level
uncorrelated wide-band noise (amplitude typically LSB/2) is less
intrusive than square wave noise.
Dithering

Original

8-colors no dither

8-colors + dither
Oversampling
• If the desired sampling rate is X, oversampling will perform the
analog-to-digital conversion at some faster rate, such as 2X.

• The technique can be used to: minimize aliasing, noise reduction and
increase accuracy beyond that provided by the wordlength.
• It widens the range of the frequency spectrum thus reducing the
(uniformly distributed) noise below the Nyquist frequency.
• When the final filtering is performed, the residual quantization noise in
the audible signal will be less: 4X oversampling yields a 6 dB
reduction (12 dB for 8X oversampling)
Storage Requirements

Type Wordlength SamplingRate SQNR Bytes/minute/chan

nel

CD 16 bits 44100 96 dB 5,292,000

CD 16 bits 48000 96 dB 5,760,000

DVD 24 bits 88200 144 dB 10,584,000

DVD 24 bits 96000 144 dB 11,520,000

DVD 24 bits 192000 144 dB 23,040,000

Storage requirement = fs * wordlength * duration * channels

DAC and Imaging
• Just as we used an ADC to go from x(t) to x(n), we can turn a
discrete sequence into a continuous voltage-level signal using a
Digital-to-analog converter (DAC).
• However, the quantized nature of the digital signal produces a
“Zero-Order Hold” effect that distorts the converted signal,
introducing some step (fast) changes.
• This distortion is know as imaging.
• To avoid this, we use a low-pass filter after the DAC, such that it
smoothes out those fast changes.
• The filter, known as an anti-imaging filter (AKA smoothing or
reconstruction filter), discards signal components above the
Nyquist frequency, thus performing a simple interpolation
between the sampled values.
Digital Recording and Playback

This is not only storage, this is

our digital system!

That system is supposed to

process the signal somehow

Still we do not know anything

about our system
Digital systems
• The digital system can be seen as an algorithm that operates on
the discrete input sequence x(n)
• The output of such a system is the sequence y(n)
• The simplest of such systems are known as Linear Time-invariant
(LTI) systems
• As the name indicates they must be time-invariant: i.e. their
behavior does not change over time; and linear: they fulfill the
following condition:
if x(n) = A " x1 (n) + B " x 2 (n)
then y(n) = A " y1 (n) + B " y 2 (n)
• For any constant A and B, and for a system where yi(n) is the
output of xi(n), thus satisfying the superposition and scaling
properties
!
Impulse response
• The input/output relations on a LTI system can be characterized
using a test signal
• A commonly-used test signal is the unit impulse, defined as:

#1 n=0
" (n) = $
%0 elsewhere
• If we apply a unit impulse to a digital system we obtain y(n) = h(n),
the impulse response of the system.
• A digital system can be completely characterized by its impulse
response
!
Discrete convolution
• Since we know the impulse response h(n) of a given system, we
can calculate its response to ANY input signal x(n) by convolving
the input with its impulse response:
m=%
y(n) = x(n) " h(n) = & x(n) # h(n $ m)
m=$%
• A convolution represents the amount of overlap between x(n) and
a reversed and temporally-shifted version of h(n)

http://mathworld.wolfram.com/Convolution.html
Basic systems
• A 2-sample delay can be described by the relation: y(n) = x(n-2)

• A gain of a is represented as: y(n) = ax(n)

• The addition (mixing) of two inputs is: y(n) = a1x2(n)+a2x2(n)

Basic systems
• By combining the previous systems we can obtain a typical digital
system:
1 1 1
y(n) = x(n) + x(n "1) + x(n " 2)
3 3 3

!
Transfer function
• However, the temporal relations between input and output are not
all we can use to describe the system
• The frequency-domain behavior of a digital system specifies which
input frequencies will be passed, rejected or emphasized.
• This behavior can be described using the transfer function H(z)
and the frequency response H(f) (that will be discussed later)
• The transfer function is obtained by calculating the Z-transform:
$
X(z) = % x(n) " z #n

n=#$

• Of the impulse response h(n) as:

$
! H(z) = % h(n) " z #n

n=#$

!
Causality and stability
• Some common Z-transforms:

x(n) X(z)
x(n " M) z"M # X(z)
$ (n) 1
$ (n " M) z"M

• Finally, to be realizable, digital systems must be:

1. Causal: the!system cannot react to an input before it is received
2. Stable: the sum of the absolute values of h(n) has to be less than
infinite
Basic Systems in MSP
• MSP is a set of extensions to Max that provide for audio analysis,
processing and synthesis
• All MSP objects end with a tilde ‘~’ to indicate audio-rate
processing. This because the tilde vaguely resembles a sine wave.

startwindow

adc~
cycle~ 440
cycle~ 440
stop
+~ turn audio *~ 0.4 multiply by number
*~ 0.2 on/off < 1.0 to attenuate
*~ 0.5 dac~
dac~

Send any discrete sequence Mix Change gain

to the DAC
Basic systems in MSP
signal in adc~

store in tapin~ 1000

delay line

read out with tapout~ 100 tapout~ 200 read out with
100 ms delay 200 ms delay
dac~

• A tapin~ object saves some amount of its input signal in a

buffer whose size is specified by the object’s argument (here
1000 milliseconds).
• Any tapout~ objects connected to the outlet of a tapin~ share
that same buffer, reading samples out after a delay.
Useful References
• Zölzer, U. (Ed). “DAFX: Digital Audio Effects”. John Wiley and Sons (2002)
• Chapter 1: Zölzer, U. “Introduction”.

• Pohlmann, K. “Principles of Digital Audio”. McGraw-Hill, Inc. (1995)

• Roads, C. “The Computer Music Tutorial”. MIT Press (1996)