Wavelet packet based features selection and fuzzy ARTMAP neural network classifier for speech classification

This paper presents an accurate voiced/unvoiced/ transition/silence speech classifier that is used as an integral part of any toll quality speech coder. Due to ability of wavelet packet in decomposition of time-frequency plane with high resolution, first five discriminate features based on energy concentration of wavelet packet coefficients for each speech classes in time-frequency plane are extracted. Then a fuzzy ARTMAP neural network classifier, which has been shown a powerful tool for non-stationary signal classification is employed. Experimental results show the proposed approach gives considerable performance improvements in some aspects with respects to the conventional methods.