[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Cochlear vs. psychoacoustic models

Dear List,

The recent discussion concerning cochlear vs. psychoacoustic models
deserves further attention, and I wish to share my experience using the
two approaches.

For my Ph.D. project (1991 - 1993) I wished to use some auditory model
as a front-end to an objective measure of sound quality, with special
focus on hearing aids.  This meant that the model besides normal hearing
should be able simulate a hearing loss, given by an audiogram.

I needed a practical "engineering" model that could take real signals,
with SPL calibration, recorded in free field, and process them into some
kind of auditory "spectrogram", i.e. excitation, firing pattern,
specific loudness etc.

The cochlear models seemed relevant with models of active outer hair
cells for increased tuning and sensitivity, and basilar membrane
mechanical model, so I started out implementing such a filter chain,
similar to ideas from Allen and Kates.  However, as soon as I wanted to
model a particular hearing loss or even make sensible estimates of the
many parameters in the model, I was helpless.  The literature on
auditory physiology does provide tuning data for animals only, who had
been impaired by a toxic or by extreme noise exposure, but this was
still a long way from an ordinary sensorineural hearing loss!  I was
also in doubt how to implement some kind of loudness summation, to come
up with sones.

So, back to the literature.  The classic psychoacoustic model is the
Zwicker school, however the procedure for deriving excitation patterns
with upward spread of masking is complicated and graphical (ISO532B),
and there is no direct information on the filter shape, let alone useful
information on how to model a hearing loss.

The answer came from Cambridge.  In several papers, Moore and Glasberg,
described their model of auditory filters, the roex-filters, WITH level
AND hearing-loss dependency WITH real numbers.  The filter bands are
ERB's (not critical bands) and the frequency scale is E (not Bark).
They even provided FORTRAN listings, in case of doubt (believe it, it's
hard to implement something based on a paper alone).

The model structure then looked like this:  Short-term FFT (I know this
could be better to avoid the fixed time resolution, i.e. some sort of
wavelet), followed by various corrections for outer ear, middle ear and
equal loudness contours, followed by a level-dependent bank of
roex-filters, whose shape also depended on hearing loss, followed by
Zwicker's loudness models to estimate specific loudness (son/ERB).  Some
other output variable could have been chosen, but specific loudness made
sense in order to include the elevated absolute thresholds.  The whole
thing was adjusted and fitted according to significant psychoacoustic
reports from the literature (in 1993).

The resulting model was able to predict common experiments well:  1 kHz
masking patterns at various levels, uniformly exciting noise, impaired
frequency selectivity, loudness growth for normal hearing and hearing
loss, and equal loudness contours (ISO 226).  There was some
underestimation of upward spread of masking and consequently also
loudness, at high levels.

Temporal properties were not simulated, nor investigated due to time
limitations in the project.

The model acted well as a pre-processor for a neural network and managed
to predict subjective ratings of sound quality by hearing-impaired and
normal-hearing listeners well.

 I think cochlear models are useful to model and study the basic
cochlear mechanisms, but far from useful to make "measuring instrument"
models.  There are newer models that can simulate a certain percentage
of inner- and outer-haircell loss along the basilar membrane, but how
does one estimate this from a regular audiogram.  Granted, the
psychoacoustic model to some extent is a "black box" simulation, but it
works in real life!

I hope I haven't generated a "flame" by writing this, but this is an
important issue to our list.  For those interested, I can provide a
literature reference list.


Lars Bramslow

Lars Bramslow
Bruel & Kjaer
Skodsborgvej 307
DK-2850 Naerum

tel (front desk) +45 45 80 05 00
tel (direct):    +45 45 80 78 55 (tone) 2663
fax:             +45 45 80 40 82

email:           lbramslow@bk.dk