Stochastic vector quantization of LPC parameters

Abstract

In this paper various issues related to efficient low‐bit‐rate coding of LPC parameters are discussed. It has been shown previously [Juang et al., IEEE Trans. Acoust. Speech Signal Process. ASSP‐30, 294–304 (1982)] that vector quantization provides a significant reduction in bit rate for coding the LPC parameters. Vector quantizers have two main disadvantages: To keep the quantization distortion to a low level, a very large codebook of vectors is usually needed. It is very difficult to train such large codebooks on real speech data and it is impractical to search these codebooks for optimum codewords in real time. In this paper, a vector quantization procedure which uses a codebook of white Gaussian random numbers to encode the LPC parameters is discussed. Interparameter correlations of LPC parameters are estimated from previously quantized data and are used to create codewords with the correct distribution. The optimum codeword is selected by an exhaustive search to minimize a weighted Euclidean distance between original and quantized parameters. Use of random codewords eliminates training of the codebook and provides robust performance over different speakers and speaking environments.