Speech compression using discreet wavelet transform

Speech compression is a process of converting human speech signals into efficient encoded representations that can be decoded back to produce a close approximation of the original signals. The paper attempts to evaluate the wavelet compression technique on speech signals. Different wavelet filters were used to select the best filter suitable for speech signals in terms of providing low bit rate and low computational complexity. We applied five procedures: one-dimensional wavelet decomposition, thresholding, quantization, Huffman coding and reconstruction using several families of wavelet filters. Our implementation was evaluated based on PSNR, SNR, NRMSE (normalized root mean square error) and compression ratio and tested on 8 kHz 8-bit speech signals. The results of this study showed that the Db10 wavelet filter gives higher SNR and better speech quality than other filters and compression ratio up to 4.31 times with satisfactory quality of decoded speech signals. In other words, the bit rate of speech signals was reduced from 64 kbps to 13 kbps.