A Speech Recognition System using an Auditory Model and TOM Neural Network

Abstract

This paper is devoted to a neurobiologically plausible approach for the design of speech processing systems. The temporal organization map (TOM) neural net model is a connectionist model for time representation. The definition of a generic neural unit, inspired by the neurobiological model of the cortical column, allows the model to be used for problems including the temporal dimension. In the framework of automatic speech recognition, TOM has been previously tested with conventional techniques of signal processing. An auditory model as front-end processor is now used with TOM, in order to test the efficiency and the accuracy of a physiologically based speech recognition system. Preliminary results axe presented for speaker-dependent and speaker-independent speech recognition experiments. The interest of auditory model is the possibility to develop more valuable processing and communication strategies between TOM and the front-end processor, including afferent and efferent information flow.