Biorthoganal Wavelet Packets and Mel Scale Analysis for Automatic Recognition of Arabic Speech via Radial Basis Functions

In this paper, a Neural Network (NN) approach for the recognition of the Arabic digits is presented. The two phases of training and testing in a Radial Basis Functions (RBF) type network is described. Biorthogonal Wavelets are constructed and used for analysis of generated subwords of the digits. This approach decomposes spoken Arabic digits based on the acoustical information contained within the speech signals. The procedure locates the boundaries between subwords by finding the peaks in the function representing the spectral changes between consecutive speech frames. The Frame-based energy parameters derived from a Wavelet Packet Scale (WPS) are used in deriving the Spectral Variation Function (SVF). Three Biorthogonal wavelets are used as analyzing functions and their performances are compared with that of their Orthogonal counterpart and with that of the traditional Fourier based Mel scale approach. Key–Words: Biorthogonal Wavelets, Radial Basis Functions, Recognizing Arabic Speech.