Cooperative modular neural predictive coding

Abstract

Speech feature extraction is one of the most important stage in the speech recognition process. In this paper, we propose a new neural networks architecture called the cooperative modular neural predictive coding (CMNPC). It is based on the interaction of discriminant experts DFE-NPC (discriminant feature extraction) optimized for macro-classification by the help of a criterion: the modelisation error ratio (MER). We propose a theoretical validation of this model by linking The MER with a likelihood ratio. The performances of this architecture are estimated in a phoneme recognition task. The phonemes are extracted from the Darpa-Timit speech database. Comparisons with coding methods (LPC, MFCC, PLP) are presented. They put in obviousness an improvement of the recognition rates.