Tuomi & Zacharov Real-time binaural loudness meter

A REAL-TIME BINAURAL LOUDNESS METER

Olli Tuomi and Nick Zacharov

Nokia Research Center, Speech and Audio Systems Laboratory,

P.O. Box 100, FIN-33721 Tampere, FINLAND
e-mail:[olli.tuomi,nick.zacharov]@nokia.com

ABSTRACT

The alignment of sound intensity is of great importance torso simulator (HATS) HRTF [10] and is
particularly in the field of psychometrics. When sounds illustrated in figure 1. It can be observed that
have different spectral content, loudness alignment there is a clear and significant directional effect,
becomes a more suitable tool. However, traditional which supports the findings of Robinson &
models assume the source either to reside in a free field Whittle [11], and Sørensen et al [12].
at 0° azimuth or in a diffuse field. To assess or align
sound source loudness at different locations or with
different radiation characteristics is more problematic.

1. INTRODUCTION 0.6
Specific Loudness (sone)

0.5

When performing psychometric studies of sound, 0.4

it is well known that normalisation of signals for 0.3

equal loudness is of critical importance. This has 0.2

often been illustrated as found by the work of 0.1

Aarts and Bech [1-4]. Quite often intensity based 0

350

alignment are suitable for simple signals, 300

250 40
200 30
however, for signals with different spectral 150
20
100
content, loudness alignment is know to be 50 10
0 0
Angle (Deg)
superior. Several models have been developed ERB

including the well known and standardised Figure 1 Simulated monaural specific loudness
Zwicker loudness model [5-7]. In more recent spectrum (left ear) of white noise as a function of
azimuth for a head and torso simulator
times, Moore has developed [8] and refined
loudness models [9]. Both are designed for use in
either the diffuse field or free field with a free According to Moore [9], it is possible to obtain an
field microphone. In the latter case, it is assumed approximation of the binaural loudness response,
that the loudness spectrum measured represents a simply by the summation of the monaural
sound source situated in front of a human head at loudness spectra. This has been simulated and is
an azimuth and elevation of 0°. However, this is illustrated in figure 2.
often not the case in practical situations
particularly when researching spatial, From these simulations, a clear need for a
multichannel or 3D reproduction sound systems. binaural loudness model is desired particularly for
Furthermore, it is rare for a real room to exhibit the loudness alignment of reproduction systems
either truly diffuse or free field characteristics. with multiple sound sources at varying locations.

In the situation where sources are in a free field at Why the need for a real-time tool? Whilst
other angles than 0° azimuth and elevation, it is alignment of sound pressure level can easily be
possible to apply a different free field to ear drum performed with dB gain compensations, the
head related transfer function (HRTF) for that mapping between the loudness and gain domain is
source angle for each ear. In this manner it non-linear and complex. By developing a real-
becomes possible to simulate the free field time loudness meter it becomes possible to
directional loudness properties for a head. This measure loudness in real-time and adjust gains for
has been performed for a one ear of a head and accurate and easy loudness alignment.

139th meeting of the ASA 1 (7)

Tuomi & Zacharov Real-time binaural loudness meter

use C-code to create the data structures in the

Matlab workspace. This allowed easy access of
the structures and verification of the data from
1 within Matlab, and greater computational
Specific loudness (sone)

0.8 efficiency.
0.6

0.4
The basic structure of the data structures had to be
altered, because all the data had to be pre-
0.2
calculated. This results in a data structure of
0
350 greater than 10 megabytes, when the original
300
250 40 model only needed a fraction of this. The memory
200 30
150
100
20 requirements would have been even greater, had
Angle (deg)
50
0 0
10
special techniques not been applied to compress
ERB
the data.
Figure 2 Simulated binaural specific loudness
spectrum of white noise as a function of azimuth for a
As a result of the above, the speed of the model
head and torso simulator
would be increased considerably. In comparison
to a version of the model implemented in pure
2. TOOL DESIGN Matlab code, the divided C-approach would run
approximately 100 times faster. The model could
be used to calculate the specific loudness of 8192
2.1. STARTING POINT sample window in 20 milliseconds. This was fast
enough to build a real-time application with the
The starting point for the tool development was model.
the Fortran version of Moore’s model, which had
been translated into a version running in the 2.3. PRACTICAL ISSUES
Matlab environment. This version was the basis
for the development of the C-language version of In the full version of the model, the biggest
the algorithm. C was selected as the latencies are no longer due to the computation,
implementation language due to the limited speed but are a result of the plotting processes. Because
of the Matlab in-built functions and structures. of this, an option to turn off the plotting was
Matlab has a means of making the use of C- implemented. This removes relevant information
language routines within native Matlab code from the user interface, but increases the
possible, called MEX routines. efficiency of the model tenfold. When the UI is
stopped, this information is updated in the UI.
2.2. BASIC TOOL STRUCTURE This makes it possible to run the model efficiently
even on a slow pentium–class machine. An
The C-code part of the model was divided into alternative approach to this would be to
two parts in an attempt to reach a more efficient implement the plotting routines in a more efficient
structure for the code. In simple implementations manner using some external graphics library, but
of the model, a great portion of time is spent in this was not done, because the required
creating the large data structures required by the processing power was available.
model. This has been avoided by taking out the
data structure creation from the computational 2.4. THE USER INTERFACE
core and placing it in a separate module.
The structure of the tool is based on the action
This approach brings in additional problems with driven UI model supported by the Matlab
memory management. As the data structures are graphical user interface generator. In the
created separately from the actual computation, in development of the user interface special care was
a different process, a way to share these structures taken to ensure easy and reliable operation. All
had to be devised. Different processes generally relevant controls are displayed in the main screen,
would need operating system calls to be able to and the user has access to all data as the
share data structures. The solution selected was to measurement is performed.

139th meeting of the ASA 2(7)

Tuomi & Zacharov Real-time binaural loudness meter

The tool offers OS independent operation with functions frame by frame, first applying a gain
sample rates up to 48kHz, 16bit samples, mono block as a way of aligning the input data to a
and stereo operation. The use of any kind of known scale. Calibration is achieved using an
signal source is supported, the model supports external calibration signal and consequently
diffuse field, free field and a special Brüel & adjusting the gain of the calibration block.
Kjær (B&K) HATS measurements mode. The
tool provides a constant display of overall Hamming windowing is applied to the data, in
loudness in sones, phons and sound pressure level preparation for the FFT algorithm. The output
in dB with plotting of specific loudness or from the FFT is excessive for input to the model,
frequency domain graphical data. Results can be so that a rescaling of the data was required prior
saved and several takes can be averaged with the to the model. The algorithm applied to create the
associated averaging tool. nonlinear scaling was specifically designed in C
for speed, large windows and different frequency
The user interface of the tool is depicted in figure grids, such as the ERB and Bark scales.
3. The main window shows the specific loudness
of the signal shown in the lower part of the The signal is then forwarded to the computational
interface. routines of the loudness model described by
Moore [9], but our specification required that an
2.5. SIGNAL FLOW additional correction mode be built into the
model. This mode would bypass any outer-ear
The signal routing through the system is presented correction, but still use the middle-ear correction
in figure 4. The signal is first routed through the block, and would be used in conjunction with the
sound hardware and operating system routines, B&K HATS.
which forward it to the data acquisition engine.
The model reads data using the DAQ toolbox

3.5

2.5

so
ne 2
s

1.5

0.5

0
0 5 10 15 20 25 30 35 40
ERBs

Figure 3 Example user interface

139th meeting of the ASA 3(7)

Tuomi & Zacharov Real-time binaural loudness meter

MATLAB DAQ Calibration Gain

SIGNAL TYPE FREQUENCIES
toolbox input domain pure sine waves 250Hz, 500Hz,
1kHz, 2kHz, 5kHz
MATLAB FFT & ERB Frq domain 2 sines, same ERB 190Hz + 200Hz
scaling output 2 sines, different ERBs 190Hz + 450Hz
OS driver &
volume controls noise pink noise
Outmidear corrected noise white noise
correction FFT plots
OS - side Table 1 Primary test signals employed for
verification
Soundcard input Moore’s specific
model loudness out In all tests, the signal was amplified using the
B&K NEXUS conditioning amplifier set to 3.16
soundcard model side V/Pa. The Windows mixing panel was used to
adjust the signal levels to maximize the dynamic
Figure 4 Model structure range of the soundcard, prior to calibration.

3. VERIFICATION TESTS All signals were generated with a DENON

professional CD player type DN-8680, and the
The HATS binaural model was to be verified by signal was routed through a YAMAHA digital
comparison of performance with respect to the mixing console type 01V.
free field microphone model under ideal
conditions in the original model. The range of Prior to the verification tests, the system gains and
conditions considered included: microphones are calibrated via an in-built
calibration routine. The free-field 4133
• Free field condition, 0° azimuth microphone was calibrated to 94dB with the
source calibrator B&K type 4231 calibrator. The 4128
• Diffuse field conditions HATS is calibrated to 97.1dB employing the UA-
• A range of test signals (See table 1.) 1546 adapter.

3.1 TEST SETUP Pure sine waves were selected to verify the basic
operation of the tool under simple signals. Signal
In order to verify the correct operation of the tool, intensities and frequencies are easy to calculate
and that it produced correct results, a series of using simple signals. White noise was employed
verification tests was planned. These tests would due to its flat spectral characteristics whilst pink
include tests involving all the functional parts of noise is more psychoacoustically motivated and
the model in different environments. broadly excites the audible bandwidth. The
combination of two sine waves in the same or
The test were conducted on a computer equipped different critical bands was selected to verify that
with a Pentium III processor of 450MHz, the basic masking properties of the signals were
128Mbytes of memory, operating system visible from the output of the tool.
Windows NT 4 service pack 5, Matlab version
R11 and using a Crystal integrated soundcard. In the test, the inputs of the 4133 and HATS are
compared. In all resulting figures, the overall
Microphones: B&K 4133 and B&K 4128 HATS. loudness is calculated for a binaural signal, but
the individual plots are the monaural responses.
Sound source: Genelec 1030A The binaural loudness in sones for monaural input
as in the case of 4133 is calculated as 2 * sones,
Artificial test signals were created with the Sonic for ease of comparison.
Foundry Sound Forge32, which were then
transferred to an audio CD to be played with a CD 3.2 FREE FIELD
player. The test signals used are shown in table 1.
The source was set up in the anechoic chamber at
2m from the microphones. The responses were

139th meeting of the ASA 4(7)

Tuomi & Zacharov Real-time binaural loudness meter

Specific loudness :1k

first measured with the calibrated B&K 4133 3.5
4133−free−m overall loudness :90.7816 phons
hats−hats−l overall loudness :92.342 phons
microphone at the effective centre of the non- 3
hats−hats−r overall loudness :92.342 phons

present HATS. The procedure was repeated for

2.5
the HATS. The results of these measurements are

Specific loudness (sone)

illustrated in figures 5-8. 2

1.5

Overall, it can be stated that there is a good

correspondence between the microphone and 1

HATS results, both in terms of the loudness 0.5

spectrum and the resulting overall loudness 0

values. However, it can be noted, particularly at 0 5 10 15 20 25

frequency (ERB)
30 35 40 45

high frequencies, that there are some differences Figure 5 Free field specific loudness spectrum
between the results. This may be cause by three for a 1kHz sine wave (0° azimuth)
factors. Firstly, the sound source is not an ideal 1.8
Specific loudness :pink

point source and thus suffers from diffraction, and 1.6

directivity associated with a two-way speaker 1.4

design. Additionally, some differences may exist

1.2

between the free field to ear drum transfer

Specific loudness (sone)

1
function employed in the original microphone and
0.8
that of the HATS. Thirdly, there is a slight
horizontal offset between the microphone position 0.6

and the entrance of the HATS ear canal (~9cm). 0.4

4133−free−m overall loudness :99.0687 phons
hats−hats−l overall loudness :99.3737 phons
0.2 hats−hats−r overall loudness :99.3737 phons

3.3 DIFFUSE FIELD 0

0 5 10 15 20 25
frequency (ERB)
30 35 40 45

Six point sources were set up in the standard Figure 6 Free field specific loudness spectrum
for pink noise (0° azimuth)
reverberation chamber to create a diffuse field
Specific loudness :190+450
into the chamber. Measurements were made in a 2
hats−hats−l overall loudness :88.4884 phons
hats−hats−r overall loudness :88.4884 phons

similar fashion to the free field case. However, to 1.8 4133−free−m overall loudness :88.5321 phons

improve the diffuse field approximation, 1.6

measurements were made in eight different 1.4

Specific loudness (sone)

1.2
positions and then results were averaged across
1
these positions. The results are illustrated in
0.8
figures 9-12.
0.6

0.4

Once again good agreement can be found between 0.2

the two sets of responses. However, the variations 0

0 5 10 15 20 25 30 35 40 45
between the two responses are greater than in the frequency (ERB)

free test, particularly at high frequencies. In Figure 7 Free field specific loudness spectrum
additions to the error sources discussed earlier, for a 190 and 450 Hz sine wave signal (0° azimuth)
which also apply here, the diffusion of the 1.5
Specific loudness :190+200

hats−hats−l overall loudness :82.8114 phons

chamber may provide some additional error. hats−hats−r overall loudness :82.8114 phons
4133−free−m overall loudness :83.64 phons

Firstly, such standard chamber are intended to be

diffuse only in a 1/3rd octave sense and with an 1
Specific loudness (sone)

upper cutoff of ~10kHz. Secondly, microphone

positional variations may also lead to some error.
The directivity of the HATS vs. the microphone
0.5
will also contribute to differences. This is
considered to be the source of variation in figure
9, which is real.
0
0 5 10 15 20 25 30 35 40 45
frequency (ERB)

Figure 8 Free field specific loudness spectrum

for a 190 and 200 Hz sine wave signal (0° azimuth)

139th meeting of the ASA 5(7)

Tuomi & Zacharov Real-time binaural loudness meter

Specific loudness :1k

2.5
4. DIRECTIONAL LOUDNESS
4133−diff−m overall loudness :82.0674 phons
hats−hats−l overall loudness :85.9392 phons
hats−hats−r overall loudness :85.9392 phons
PERFORMANCE
2

Having established that the performance of the

Specific loudness (sone)

1.5
binaural implementation is in line with the
common and standard models, it is now of interest
1
to verify the directional loudness performance
compared to simulations presented earlier.
0.5

0
To achieve this the HATS was mounted on an
0 5 10 15 20 25
frequency (ERB)
30 35 40 45
automated turntable. White noise was reproduced
Figure 9 Diffuse field specific loudness by a source at 2m distance in an anechoic
spectrum for a 1kHz sine wave (diffuse field) chamber and the loudness spectrum data was
1
Specific loudness : pink collected in the horizontal plane in 10°
0.9
increments. The monaural results are illustrated in
0.8
figure 13 below and can be directly compared to
0.7 the simulated version found in figure 1. The
binaural measurements are shown in figure 14 and
Specific loudness (sone)

0.6

0.5 compare favorably to simulations shown in figure

0.4 2. A clear similarity can be found in both cases.
0.3

0.2 4133−diff−m overall loudness :94.5571 phons

hats−hats−l overall loudness :94.9969 phons
hats−hats−r overall loudness :94.9969 phons
0.1 1.4

0 1.2
0 5 10 15 20 25 30 35 40 45
frequency (ERB)
Specific loudness (sone)

Figure 10 Diffuse field specific loudness 0.8

spectrum for pink noise (measured in a diffuse field)
0.6
Specific loudness :190+450
3 0.4
hats−hats−l overall loudness :91.0324 phons
hats−hats−r overall loudness :91.0324 phons
4133−diff−m overall loudness :91.6282 phons 0.2
2.5
0
400

2 300 40
Specific loudness (sone)

30
200
20
1.5 100
10
0 0
Angle (deg)
Frequency (ERB)
1

Figure 13 Measured monaural specific loudness

0.5
spectrum (left ear) of white noise as a function of
azimuth for the B&K 4128 HATS
0
0 5 10 15 20 25 30 35 40 45
frequency (ERB)

Figure 11 Diffuse field specific loudness

spectrum for a 190 and 450 Hz sine wave signal 2.5

Specific loudness :190+200

3
2
Specific loudness (sone)

4133−diff−m overall loudness :90.5877 phons

hats−hats−l overall loudness :89.1423 phons
hats−hats−r overall loudness :89.1423 phons
1.5
2.5

1
2
Specific loudness (sone)

0.5

1.5
0
400

1 300 40
30
200
20
0.5 100
10
0 0
Angle (deg)
Frequency (ERB)
0
0 5 10 15 20 25 30 35 40 45
frequency (ERB) Figure 14 Measured binaural specific loudness
Figure 12 Diffuse field specific loudness spectrum (left ear) of white noise as a function of
spectrum for a 190 and 200 Hz sine wave signal azimuth for the B&K 4128 HATS

139th meeting of the ASA 6(7)

Tuomi & Zacharov Real-time binaural loudness meter

These tools have been employed for the loudness of the Audio Engineering Society, Vol. 39,
alignment of multiple sound reproduction systems pp.27-38, January/February 1991.
with different number of sources in a standard [2] Aarts R. M., Comparison of Some Loudness
listening room. Mono, stereo and multichannel Measures for Loudspeaker Listening Tests,
systems were loudness aligned with this tool and Journal of the Audio Engineering Society,
the informal subjective alignment was found to be
Vol. 40, pp.142-146, March 1992.
very satisfactory.
[3] Aarts R. M., On the design and
psychophysical assessment of loudspeaker
5. SUMMARY AND CONCLUSIONS
system, Proefschrift Technische Universiteit
A real-time binaural loudness meter has been Delft, 1995.
efficiently implemented for use in the Windows [4] Bech, S., Calibration of relative level
PC environment. The system employs the systems differences of a domestic multichannel sound
native soundcard for use with measurement reproduction system, J. Audio Eng. Soc., vol.
microphones and allows for calibration of all 46, pp 304 – 313, April 1998.
gains within the input chain. The performance has [5] Paulus E., Zwicker E., Programme zur
been shown to be in line with the original automatischen Bestimmung der Lautheit aus
loudness model, though with real-time Terzpegeln oder Frequenzgruppenpegeln,
performance. Directional loudness properties have Acustica, Vol. 27. pp. 253-266, 1972.
also been shown to compare favorably with the [6] Zwicker E., Fastl H., Psychoacoustics, Facts
simulated data.
and Models, Springer-Verlag, 1990.
[7] ISO Rec. R. 532, Method for calculating
6. FUTURE WORK
Loudness Level, Method B, International
Future work with the model includes creating a Organization for Standardization, Geneva,
standalone version of the tool, that could be used Switzerland, 1966.
without the Matlab environment, greatly lowering [8] Moore B. C. J., Glasberg B. R., A revision of
the price of the tool. As the Zwicker model is still Zwicker’s Loudness Model, Acustica,
in widespread use in the audio engineering Vol.82, pp. 335-345, 1996.
community, it will be implemented into the model [9] Moore B. C. J., Glasberg B. R., and Baer T.,
as an alternative. A model for the prediction of thresholds,
loudness, and partial loudness, J. Audio Eng.
If the tool is to be used with transient signals, Soc., vol. 45, pp 224 – 239, April 1997.
triggering will have to be implemented. [10] N. Zacharov, An overview of multichannel
Furthermore, a better way of approximating
level alignment, in Proc. of the AES 15th Int.
temporal loudness integration is required, as
Conf., pp. 174–186, Audio Eng. Soc., 1998.
proposed in [13].
[11] Robinson, D. W., Whittle, L. S., The
loudness of directional sound fields,
7. ACKNOWLEDGEMENTS
Acustica, vol 10, (1960), pp 74 – 80
Prof. Brian Moore is thanked for his comments [12] Sørensen M. F., Lydolf M., Frandsen P. C.,
and assistance in the development of this tool. Møller H., Directional dependence of
Pekka Suokuisma and Miikka Vilermo are loudness cues and binaural summation,
thanked for their assistance with implementations. proceedings of the 15th international
Th second author expresses thanks to the Nokia congress on acoustics, Trondheim, Norway,
Foundation for funding support during finalising pp. 293-296, June 1995.
this study. [13] Stone, M. A., Moore, B. C. J., and Glasberg,
B. R., A real-time DSP based loudness meter,
8. REFERENCES Seventh Oldenburg symposium on
psychological acoustics, Germany, pp. 587-
[1] Aarts R. M., Calculation of the Loudness of 601, 1997.
Loudspeakers during Listening Tests, Journal

139th meeting of the ASA 7(7)