Incorporating perception into LSF quantization some experiments

In the context of vector quantization (VQ) of the line spectrum frequency (LSF) parameters, we determine experimentally a spectral distribution of quantization error perceived to be " balanced " , i.e., error at all frequencies contributing equally, on average, to the perceived distortion. Quantizers which have a balanced distribution should outperform those which don't, given the same number of bits. We examine the spectral error distributions produced by various weighted Euclidean distance measures in the LSF domain and develop one which produces a quantizer having an approximately balanced distribution. This quantizer's performance is compared with that of others having different error distributions. 1. INTRODUCTION The short term speech spectrum contains important information and is therefore quantized and transmitted by many speech coding algorithms. This information is often determined by linear prediction and can be represented by many spectral parameter sets such as predictor coefficients, log area ratios, reflection coefficients, LSF's, etc. Accurate representation of the spectrum using as few bits as possible is desired. Prior research has shown that VQ of the LSF parameters is a relatively efficient method for minimizing the number of bits needed to represent the speech spectrum. An example is the new 2400 bps MELP coder which uses a 25 bit, multi-stage vector quantizer (MSVQ) to encode 10 LSF coefficients [1]. While LSF VQ's perform well enough to be used in real-time coders, numerous aspects of their design and evaluation could benefit from additional research. One of these is how to make better use of human perceptual characteristics which cause decreased sensitivity to spectral error as the error frequency increases. Some VQ work has used this property in a quantizer performance measure [1][2]. However, we are unaware of any research to determine the amount of spectral error needed for audi-bility as a function of frequency. This paper describes experiments which investigate relationships among spectral error, perception, and weighted Euclidean distance measures. We determine experimentally a spectral error distribution that is approximately " balanced " , i.e., error at all frequencies