A Computational Model of Auditory Feature Extraction and Sound Classification

Abstract

This thesis introduces a computer model that incorporates responses similar to those found in the cochlea, in sub-corticai auditory processing, and in auditory cortex. The principle aim of this work is to show that this can form the basis for a biologically plausible mechanism of auditory stimulus classification. We will show that this classification is robust to stimulus variation and time compression. In addition, the response of the system is shown to support multiple, concurrent, behaviourally relevant classifications of natural stimuli (speech). The model incorporates transient enhancement, an ensemble of spectro - temporal filters, and a simple measure analogous to the idea of visual salience to produce a quasi-static description of the stimulus suitable either for classification with an analogue artificial neural network or, using appropriate rate coding, a classifier based on artificial spiking neurons. We also show that the spectotemporal ensemble can be derived from a limited class of 'formative' stimuli, consistent with a developmental interpretation of ensemble formation. In addition, ensembles chosen on information theoretic grounds consist of filters with relatively simple geometries, which is consistent with reports of responses in mammalian thalamus and auditory cortex. A powerful feature of this approach is that the ensemble response, from which salient auditory events are identified, amounts to stimulus-ensemble driven method of segmentation which respects the envelope of the stimulus, and leads to a quasi-static representation of auditory events which is suitable for spike rate coding. We also present evidence that the encoded auditory events may form the basis of a representation-of-similarity, or second order isomorphism, which implies a representational space that respects similarity relationships between stimuli including novel stimuli.