AUDITORY MODELLING FOR AUTOMATIC SPEECH RECOGNITION
SW BEET, Roger K. Moore, MJ TOMLINSON · 2024
Previous work at 3.5.8.3.has shown that the resolution offered by conventional filter-bank analysis is neither sufficient for accurate temporal modelling of speech signals nor for reliable separation of speech from competing signals and noise.In order to further investigate these problems.a software model of the human peripheral auditory system has been implemented.based on that described by R.P. Lyon [1. 2. 3].Lyon's model.however.produces data at an extremely high rate and it is consequently unsuitable for immediate use by current recognition algorithms.This paper describes a reduced auditory representation (R.A.R.).which preserves the main characteristics observed in the output of the auditory model.including the ability to resolve fine temporal and spectral detail (without excessive disruption by individual pitch pulses).but which produces data at a more acceptable rate. 11-13 LYON AUDITORY MODE.Many structures have been suggested for simulating the observed behaviour of the human auditory system.but one of the most comprehensive versions to retain computational tractability is that developed by Richard Lyon at Schlumberger Palo Alto Research.This model approximates the behaviour of the inner ear.while effects due to the outer and middle ears are largely neglected.The latter are omitted from the model because they merely impose a fixed.albeit position and environment dependent.pectral distortion on any signal and it is