Classification of voiceless plosives using wavelet packet based approaches

There are contradictory reports on the usefulness of the Wavelet Packet Transform for feature extraction. In this paper we continue the investigation of this subject with reference to non-stationary speech signals, namely unvoiced plosive consonants /p/,/t/, /k/. We concentrate on the influence of the feature reduction method on the classification rate. Two strategies have been applied: feature selection, performed using the Local Discrimination Basis and feature projection performed using Primary Components Analysis (Singular Value Decomposition). Classification has been performed by cluster analysis and neural network. The classification results obtained for PCA outperform those for LDB and other methods examined earlier.