Edinburgh Napier University

  Speech feature extraction methods are commonly based on time and frequency processing approaches. In this paper, we propose a new framework based on sub-band processing and non-linear prediction. The key idea is to pre-process the speech signal by a filter bank. From the resulting signals, non-linear predictors are computed. The feature extraction method involves the association of different Neural Predictive Coding (NPC) models. We apply this new framework to phoneme classification and experiments carried out with the NTIMIT database show an improvement of the classification rates in comparison with the full-band approach. The new method is also shown to give better performance than the traditional Linear Predictive Coding (LPC), Mel Frequency Cepstral Coding (MFCC) and Perceptual Linear Prediction (PLP) methods.