Cochlear Linear Predictive Coding: Interfacing an Analogue Cochlear Model with a Conventional Speech Recognition System

Conventional speech recognition systems require as input a sequence of acoustic vectors coding the relevant features of the speech signal. On the other hand the biological cochlea (inner ear) processes the speech signal to extract a neural coding suited for the higher levels of the brain. The background idea of this article is to interface an analogue electronic model of the cochlea with a statistical classifier based on HMM. Combined with an analogue gradient descent circuit, such an artificial cochlea can be used to extract the Linear Predictive Coding (LPC) of the speech signal in continuous time. We thus propose an analogue VLSI circuit implementing a so-called 'Cochlear LPC' (CLPC). Speech recognition results obtained using a computer model of the CLPC circuit combined with an HMM classifier are presented, which compare favourably with those obtained from a standard LPC/HMM system. The advantages of the CLPC, such as its optimal time-frequency resolution and dynamics properties, as well as the continuous time processing in general are discussed. Finally, we briefly outline alternative pre-processing options based on the artificial cochlea which are expected to increase the functional compatibility with a statistical classifier.