Neural Expert Model Applied to Phonemes Recognition

Abstract

Connectionist models often offer good performance in pattern recognition and generalization, and present such qualities as natural learning ability, noise tolerance and graceful degradation. By contrast, symbolic models often present a complementary profile: they offer good performance in reasoning and deduction, and present such qualities as natural symbolic manipulation and explanation abilities. In the context of this paper, we address two limitations of artificial neural networks: the lack of explicit knowledge and the absence of temporal aspect in their implementation. STN : is a model of a specialized temporal neuron which includes both symbolic and temporal aspects. To illustrate the STN utility, we consider a system for phoneme recognition.