Non linear and discriminant feature extraction applied to phonemes recognition

In this article, we propose to study a speech coding method applied to the recognition of phonemes. The proposed model (the Neural Predictive Coding, NPC) and its two declinations (NPC-2 and DFE-NPC) is a connectionist model (multilayer perceptron) based on the non linear prediction of the speech signal. We show that it is possible to improve the discriminant capacities of such an encoder with the introduction of signal membership class information as from the coding stage. As such, it fits in with the category of DFE encoders (Discriminant Features Extraction) already proposed in literature. In this study we present a theoretical validation of the model in the hypothesis of unnoised signals and gaussian noised signals, NPC performances are compared to that obtained with traditional methods used to process speech on the Darpa Timit an Ntimit speech bases. Simulations presented here show that the classification rates are clearly improved compared to usual methods, in particular regarding phonemes considered difficult to process. A small vocabulary word recognition experiment is provided to show how NPC features can be used in a more conventional speech ANN-HMM based system approach.