Subject: Re: Cochlear vs. psychoacoustic models From: at <meijerNATLAB.RESEARCH.PHILIPS.COM> Date: Wed, 18 Jun 1997 15:50:37 +0200June 18, 1997 Dear Lars, I agree with much of what you say, and like to elaborate a little on it. Indeed there will always be good reasons for having different types of models - both cochlear and psychoacoustic, both ``physical'' and ``functional.'' In fact, it all depends on the purpose one has in mind to use a model for. If the target is a detailed understanding of low-level auditory mechanisms, then go for the physical route, applying brute-force computing to solve the many (partial) differential equations. If the target is efficient simulation, e.g., needed to pull through many complex sounds at or near the (hearing) system level, then opt for a functional model, even a ``black-box'' one if it is too hard to derive a simplified but still accurate functional model from the physical one (and quite often it *is* too hard). An advantage of functional models is also that it is often easier to switch off particular effects, allowing one to trace which modelled non-ideality causes a certain observed (functional) effect at the macroscopic level. With physical models it is often far more difficult to create, say, a given, measured, nonlinear distortion level: functional models are sometimes even more accurate than physical model in representing observable effects, because many parameters in a physical model are often only approximately known, while in a functional model distortion itself may be a parameter. The opposite can also happen, of course, with (too) simple functional models being (much) less accurate than the physical ones. Furthermore, functional models may have poor predictive value for ``new'' auditory phenomena: they are better for application areas in which the range of relevant auditory phenomena is already known/spanned. In other words, functional models can be quite good at ``interpolation'' of data sets, but are often poor at ``extrapolation.'' Another issue is that a model alone is not much good. Models need to be provided with testbenches, i.e., data sets to allow comparison of results from different models, (in)validation of model outcomes for functional versus physical models, etc. Such testbenches must obviously be mutually compatible: a single testbench must be usable with a variety of models. In my recent draft experiments with the AIM model at http://ourworld.compuserve.com/homepages/Peter_Meijer/aumodel.htm I used a .wav file that I would like to see processed by both physical and functional models of different origins in order to compare the results. Of course this data set should have been much larger, e.g., with samples that cover specific effects of temporal/frequency masking. Once the physical and functional models give similar outcomes under a wide variety of data sets (sound waves), not just few example cases - or they must be very carefully selected, confidence will grow that each type of model is good for its particular target area (e.g., understanding versus efficiency). I haven't seen a systematic approach to this so far. [A moment ago the posting from John Beerends came in while I was finishing this text. I think his detailed comments are consistent with what I remarked above about the possible high accuracy of functional models.] Best wishes, Peter Meijer