Learning Vector Quantization and Neural Predictive Coding for nonlinear speech feature extraction

Speech recognition is a special field of pattern recognition. In order to improve the performances of the systems, one can opt for several ways and among them the design of a feature extractor. This paper presents a new nonlinear feature extraction method based on the Learning Vector Quantization (LVQ) and the Neural Predictive Coding (NPC). The key idea of this work is to design a feature extractor, the NPC, by the introduction of discriminant constraint provided by the LVQ classifier. The performances are estimated on a phoneme classification task by several methods: GMM, MLP, LVQ. The phonemes are extracted from the NTIMIT database. We make comparisons with linear and nonlinear feature transformation methods (LDA, PCA, NLDA, NPCA), and also with coding methods (LPC, MFCC, PLP).