Point process models of spectro-temporal modulation events for speech recognition

Abstract

Neurobiological research has uncovered the existence of cortical neurons in various animal species tuned to particular spectrotemporal modulations (STM) in the auditory stimulus. Other findings indicate that temporal statistics of the resulting neural spike trains may encode the underlying content of species-specific communication calls. With this motivation, we present an alternative approach to speech recognition based on point process statistical models of the local maxima events produced by a cortically-inspired spectro-temporal filter bank. We demonstrate the computational adequacy of this approach on the practical task of keyword spotting.