Scalable speech coding based on the wavelet transform

This paper describes a scalable speech coder for operation at bit rates between 4 and 32 kbit/s. The bit rate can be changed with a minimum step size of one bit per frame. The algorithm is based on a predictive coding scheme using adaptive transform coding (ATC) of the residual signal based on a 20-ms frame. The shape of the excitation signal is coded open-loop while gain quantisation and long-term prediction are performed in a closed-loop way. Two different transforms, the discrete cosine transform (DCT) and the discrete wavelet transform (DWT) are investigated and compared with each other. It is shown that the DWT is superior to the DCT at low bit rates. The algorithm has been tested at various bit rates and has been compared with standardised speech coders using ITU-T Rec. P.861 (PSQM) as an objective quality measure. It turns out that the speech quality of the proposed coder is roughly equivalent to ITU-T G.729 at 12 kbit/s and ITU-T G.728 at 16 kbit/s. Even at 4 kbit/s the algorithm produces good-quality output speech.