New wavelet packet model for automatic speech recognition system

This paper introduces an automatic speaker-independent speech recognition system. We investigate the performance of the wavelet packet in the analysis of automatically generated subwords of single digits. The modeling of the subwords is accomplished using multienergy levels of a derived mel-like scale. A radial basis function artificial neural network (RBF-ANN) is employed for the recognition task. The proposed model is compared with two systems, one uses manual segmentation, the other segments words based on energy levels extracted from a filter bank. A comparison is made between the performance of systems using two orthogonal wavelets from the Daubechies set and two biorthogonal wavelets.