Optimal transform coding for speech line spectrum pair parameters based on spectral-weighted error criterion

Abstract

The line spectrum pair (LSP) is one of the most popular and efficient parameters for representing the short-time spectrum of speech signal. About 34 bits/frame is needed for direct scalar quantization of LSP parameters to maintain a good quality. Based on the spectral-weighted Euclidean distance of LSP parameters, a modified transform coding of LSP parameters which takes advantage of both strong interframe and intraframe correlations is proposed in this paper. This method incorporated with the DPCM scheme does not introduce further buffering delay. It utilizes the spectral-weighted LSP parameters in transform coding instead of LSP vector itself and needs only to quantize the prediction residuals in transformed domain. The much smaller dynamic ranges and symmetrically bell-shaped distributions of those residuals explain the success of this scheme. With the frame period of 10 ms, the spectral distortion limen of 1 dB2 can be reached at 18 bits/frame with inside training test data and 19 bits/frame with outside training test data.