Note on the Properties of a Vector Quantizer for LPC Coefficients

Note on the Properties of a Vector Quantizer for LPC Coefficients By L. R. RABINER,* M. M. SONDHI,* and S. E. LEVINSON* (Manuscript received March 15, 1983) Vector quantization has been used in coding applications for several years. Recently, quantization of linear predictive coding (LPC) vectors has been used for speech coding and recognition. In these latter applications, the only method that has been used for deriving the vector quantizer code book from a set of training vectors is the one described by Linde, Buzo, and Gray. In this paper, we compare this algorithm to several alternative algorithms and also study the properties of the resulting code books. Our conclusion is that the various algorithms that we tried gave essentially identical code books. I. INTRODUCTION The technique of vector quantization for LPC voice coding has been in use for several years, and has been shown to be of great utility for LPC analysis/synthesis systems. 1-4 Recently, vector quantization of LPC vectors has been applied to speech-recognition systems both in direct applications 5,6 and in conjunction with work on the application of hidden Markov models (HMMs) to recognition.7,8 The main idea of vector quantization is summarized as follows: assume that a training set [T of I LPC vectors is given. It is desired to find a code book of M* LPC vectors such that the average distance of a vector in {T from the closest code book entry is minimized. Thus we wish to find a set R] of reference vectors that minimizes the * Bell Laboratories.