Wavelet statistical model of speech for feature extraction and denoising

To improve the performance of automatic speech recognition (ASR) systems, a new method is introduced to extract features and reduce noise. Robust features are obtained from a wavelet packet transform (WPT)/local discriminant bases (LDB) stage, which can efficiently classify different speech utterances. The enhancement is performed by means of a feed-forward subsystem linked to the WPT/LDB stage, an ARMA/wavelet based discriminant function minimum (DFM) working as a blind adaptive filter (BAF), and an unvoiced speech enhancement stage. It also makes use of minimum description length (MDL) to reduce the number of wavelet coefficients for automatic recognition and re-synthesis. Simulation results showed that this new system is an efficient classifier and improves the robustness of ASR systems in various adverse noisy conditions.