Low bit rate line spectrum pair (LSP) quantizer with application to speech coding

This paper proposes a new efficient line spectrum pair (LSP) quantizer, which is based on the combination of the difference LSP techniques and moving average (MA) interframe prediction. The difference technique is used to remove the correlation between adjacent LSP coefficients in the present frame, while the moving average (MA) prediction removes the correlation between the corresponding coefficients in successive frames. Two techniques are used for residual quantization, the first is a low complexity scalar quantizer with 20-25 bits per frame with 20 ms frame size. The second is of higher complexity but lower bit rate which uses 14 bits per frame two stages vector quantizer (VQ). The DMAP quantization technique yields a good performance comparable to the moving average two-stage split vector quantizer used in the ITU-T G.729 standard.