Phoneme recognition with binaural cochlear models and the stereausis representation

The authors describe phoneme recognition experiments which use biologically motivated dual channel collection and processing of speech corrupted by noise. The signals are processed by head-related transfer functions to produce right and left side signals which are then used as input to cochlear models. The cochlear outputs are combined using the stereausis algorithm to produce feature vectors which are classified by a self-organizing feature map. Binaural collection and processing are shown to improve recognition rates over a single channel system for certain conditions and broad class speech sounds.<>