Discriminative training for neural predictive coding applied to speech features extraction

We present a predictive neural network called neural predictive coding (NPC). This model is used for nonlinear discriminant features extraction applied to phoneme recognition. We validate the nonlinear prediction improvement of the NPC model. We also, present a new extension of the NPC model: NPC-3. In order to evaluate the performances of the NPC-3 model, we carried out a study of Darpa-Timit phonemes (in particular /b/, /d/, /g/ and /p/, /t/, /q/ phonemes) recognition. Comparisons with traditional coding methods are presented. We also show how an adaptative constraint allows improvements on the recognition task