Vector predictive quantization of the LPC spectral parameters for low‐rate speech coding

TLDR: In this article, the use of vector predictive quantization (VPQ) for the LPC parameters is proposed, where the suboptimality is compensated for by exploiting the temporal redundancy in the input.

Abstract: Scalar quantization of the LPC parameters requires a high bit rate. Considerably lower rate can be obtained via vector quantization (VQ). However, complexity constraints dictate the use of suboptimal VQ. The use of vector predictive quantization (VPQ) for the LPC parameters is proposed. In VPQ, the suboptimality is compensated for by exploiting the temporal redundancy in the input. VPQ is a two‐stage memory VQ. In the first stage, the input vector is predicted from quantized past vectors, using a set of vector coefficients, held in a predictor codebook. In the second stage, the predicted vector is combined with a residual vector to form the final output. A set of residual vectors is held in a residual codebook. VPQ was applied to the LPC parameters in the down‐sampled, log‐magnitude spectral domain. The idea was to design an efficient VPQ under the perceptually meaningful log‐likelihood distortion measure, while circumventing the stability problem of the synthesis LPC filter. This VPQ was used in a 4.8 kbit/s CELP coder where only 1.0 kbit/s were allocated to the parameters. The performance was almost indistinguishable from that of a CELP coder with unquantized parameters. [Work supported by NSA.]