The Karhunen-Loeve transform applied to the log area ratios of a linear predictive speech coder

One problem of interest to digital speech compression researchers is to reduce the number of bits required to adequately describe the spectral information provided by linear predictive analysis of voiced speech. The set of log area ratios has been shown to be one of the best coefficient sets when quantization considerations are concerned. In this paper an optimal scheme for quantizing the log area ratios is derived based on a mean-square spectral distortion criteron. The quantization of the log area ratios is compared to the quantization of the set of coefficients resulting from the Karhunen-Loeve transform of the log area ratios. Experimental data is presented which indicates that little, if any, bit rate reduction is obtained through use of this transform.