Psychoacoustic testing of Searle's model based on human audition

Abstract

A software model was built to investigate the feasibility of reducing the dimensionality of the representation of speech for speech recognition. The model was based on a model of human audition developed by Searle in hardware, the characteristics of which closely match physiological characteristics obtained experimentally from the human auditory system. The model consisted of 16, one third octave bandpass filters followed by envelope detectors, the outputs of which were then subjected to the linear discrete cosine transform in an attempt to reduce the data to a small number of perceptually important dimensions. Several signal processing techniques were investigated to reconstruct the filtered, detected and transformed speech to permit effective intelligibility testing. Most of the work done was qualitative and the methodology concentrated on testing the model rather than the theory. A test sentence produced by a male and female speaker was transformed using the model and the transformed speech was reconstructed. Informal testing with a few experienced listeners suggests that it might be possible to recognize speech with as few as three out of the sixteen channels. This study confirms the value of realizing the models in software. Although the model is not real time it was far easier to modify than the comparable hardware models.