Wavelet packets based features selection for voiceless plosives classification

There are contradictory reports on the usefulness of the wavelet packet transform (WPT) for feature extraction. This is mainly the case of signals of non-stationary character. In this paper we examine this tool for a category of short non-stationary speech signals, namely voiceless plosive consonants /p/, /t/, /k/. Three approaches to feature selection have been implemented: best basis search algorithm over the averaged wavelet packet coefficients of all data, local discriminant basis (LDB) algorithm, i.e. application of the best basis algorithm on the discriminant measure between coefficients in three classes and singular value decomposition (SVD) of the entropy matrices calculated from the wavelet packets for each class. The experiments conducted over the context independent plosives from speech database of Polish gave a classification rate higher for WPT based features than for traditional DFT based cepstral coefficients.