Subject: AIMR8.2 From: Roy Patterson <roy.patterson(at)MRC-APU.CAM.AC.UK> Date: Mon, 7 Jul 1997 18:19:20 +0100AIMR8.2 AUDITORY IMAGE MODEL: RELEASE 8.2 May 1997 In October 1995, we published a short paper on the Auditory Image Model (AIM) and the software package that we use to run the model. Patterson, R.D., Allerhand, M., and Giguere, C., (1995). "Time-domain modelling of peripheral auditory processing: A modular architecture and a software platform," J. Acoust. Soc. Am. 98, 1890-1894. AIM converts a digitised sound into a multi-channel neural activity pattern (NAP) like that produced by the cochlea in reponse to the sound. Then it applies strobed temporal integration or autocorrelation to each channel of the NAP to convert it into something more like the auditory image we hear when presented with the sound. You can read about AIM R8 on our WWW page, and pick up the sources there, http://www.mrc-apu.cam.ac.uk/aim/ or pick up the source code and documentation by annonymous ftp from ftp.mrc-apu.cam.ac.uk directory pub/aim The ReadMe.First file explains how to to acquire the source code, how to compile it, and how to get started. The paper describes AIM Release 7.1 (AIM R7). The purpose of this letter is to announce the first public release of AIM R8 (Release 8.2, May 1997). There was also a beta version of AIMR8 (R8.1 dated August 96). There are three changes to AIM itself: a) We have introducted a bank of low-pass filters at the output of the cochlea simulation to improve the simulation of loss of phase locking at high frequencies. b) We have introduced the option of power compression at the output of the gammatone auditory filterbank. c) We have changed the nonlinearity of the transmission-line filterbank to a square-root function. We have also developed a Matlab interface for AIM, and written a document describing how to use the AIM software to create a Meddis and Hewitt (1991) model of pitch perception (docs/aimMeddisHewitt). ======================================================================= CHANGES TO THE AUDITORY IMAGE MODEL ITSELF The software package contains two basic types of AIM: functional and physiological. The modifications are as follows: 1a. Power Compression Option: compress=0.5 The compression in the functional version of AIMR7 is logarithmic and this is the appropriate function to bridge between gammatone filters with their exponential tails and two-dimensional thesholding -- the process that simulates neural transduction in functional AIM. There are now data to indicate, however, that a power compressor with an exponent of 0.5 (applied to amplitude values) is a better representation of auditory compression over quite a wide dynamic range. Accordingly, the 'compress' option has been modified to accept the arguments 'log', 'off', or a value between 0 and 1 indicating the degree of power compression required. The default is 'log'. The file bin/aimR8demo contains a script that shows how to set up functional AIM with square root compression rather than logarithmic compression. 1b. Nonlinearity of the transmission line filterbank The basilar membrane nonlinearity of the physiological version has been changed from an inverse function, 1/(1+x), to a square-root function, sqrt(1/(1+x)), to improve the simulation of auditory nonlinearities and two-tone suppression effects. The change in nonlinearity has virtually no effect on the main travelling wave for a wide range of input levels. There are no changes to the relevant options (with suffix _tlf). 2. Loss of Phase Locking: stages_idt=2 tup_idt=0.133ms In the functional version of AIM, the instruction to simulate the neural activity pattern produced by a sound (gennap) has been modified to improve the simulation of 'loss of phase locking' as a function of frequency in the region above 1200 Hz. AIMR8 applies a two-stage, or second order, lowpass filter with a time constant of 0.133 ms to the output of each NAP channel. The filtering results in a 12 dB/ocatve rolloff in temporal resolution in the region above 1200 Hz. Previously there was no loss of phase locking. The relevant options and defaults are 'stages_idt=2' and tup_idt=0.133ms'. Set 'stages_idt=off' to revert to AIMR7. The lowpass filter can also be used with the Meddis haircell module in the physiological version of AIM to increase the rate of phase locking from its original 6 dB per octave above 1050 Hz, to a more realistic 12 dB/octave above about 1100 Hz. The relevant options and defaults are 'stages_idt=1' and tup_idt=0.133ms'. The issues are described in docs/aimMeddisHewitt. 3. bits_wave: The options 'bits_wave' has been removed since it has no effect in the floating point version of AIM. It remains a silent option for anyone who still uses the integer version. 4. makefile: The makefile has been organised and comments have been added to assist compilation at non-standard unix sites. There is also a help option for the makefile ('make help'). ======================================================================= ADDITIONS TO THE SOFTWARE PACKAGE 1. Simulation of Meddis and Hewitt (1991) using the AIM software: The functional version of AIM includes a gammatone auditory filterbank. The physiological version includes a bank of Meddis haircells and a bank of autocorrelators for constructing autocorrelograms. You can cross connect the modules and so produce a Meddis and Hewitt (1991) model with a gammatone auditory filterbank, a bank of Meddis haircells and a bank of autocorrelators. The option settings required to enable the Meddis and Hewitt model are described in docs/aimMeddisHewitt. The file docs/aimR8demo contains a script that demonstrates these new versions of AIM. 2. Matlab/AIM: There is now a Matlab interface for AIM so that you can construct sounds using matlab functions, pass the waves to AIM for processing, and import the multi-channel neural activity patterns, or auditory images, back into Matlab for futher processing. It is referred to as AIMMAT, and it is particularly useful for developing decision statistics to relate the complex patterns observed in the neural patterns and auditory images to data from experiments with human listeners. It is described in Tsuzaki, M. and Patterson, R.D. (1997). "AIM and AIMMAT as simulators of auditory peripheral processing," Int. Symp on Simulation, Visualization and Auralization. 2-4 April, Tokyo, Japan. The interface tools are found in directory 'matlab'. There are a set of demonstration routines with the prefix 'amd_'. They are all matlab scripts (that is, ".m" files) and so the easiest way to get going in Matlab/AIM is to find the amd_xxx.m demo that is closest to your own interest and modify a copy of it for your own application. 3. Silent Options for Postscript printing: There are now a large number of Silent Options associated with printing AIM displays (see docs/aimSilentOptions). They are particularly useful when printing autocorrelograms, summary autocorrelograms, and summary auditory images, where some of the default axes and labels are incorrect. Examples of how to use these silent options are presented in docs/aimR8demo. 4. The Bibliography of AIM related papers (docs/aimBibilography) has been updated. 5. A new TARGET has been added to the makefile to assist compilation on SGI machines. 6. A document has been drafted describing the strobed temporal integration (STI) mechanism used to construct auditory images from neural activity patterns, and the relationship between STI and autocorrelation. See docs/aimStrobeCriterion. It is still under development but is the best description of the issues as we think of them currently. ======================================================================= Regards from the AIM team, Roy Patterson, Chris Giguere, Minoru Tsuzaki, Michael Akeroyd, Jay Datta and Mike Allerhand May 1997 =======================================================================