The line spectrum pair (LSP) is one of the most effective representations of the speech short-time spectrum. About 34 bits/frame is needed for direct quantization of LSP parameters to maintain a reasonable accuracy. Based on the spectral-sensitivity-weighted Euclidean distance of LSP parameters, a hybrid TC/DPCM coding of LSP parameters which takes into account the spectral sensitivity weighting is proposed . In addition to the scalar quantization, two vector quantization methods which are multi-stage VQ and partitioned VQ, are considered. With the frame period of 10 ms, the best result shows that the spectral distortion limen of 1 dB2 can be achieved at 18 bits/frame with in-training test data and 20 bits/frame with out-of-training test data, respectively.
Keywords: line spectrum pair, spectral distortion, vector quantization.