Optimal vector quantization of speech spectral principal components

Abstract

The vector quantization method described by Linde et al. [IEEE Trans. Commun. 28, 84–95 (1980)] was used to optimally quantize speech spectral principal components. The spectral principal components, obtained from linear combinations of spectral band energies, are an optimal uncorrelated parameter set containing a very large amount of spectral variance with a relatively small number of dimensions. For a given percentage of quantization distortion, the vector quantizer required fewer bits per frame than individual quantization of principal components, thus indicating a clustering of the speech data in the uncorrelated parameter space. Using the principal components parameters to measure both gain and spectral shape and using eight bits per frame, the mean square vector quantization error was about 2.5% in a four‐dimensional space, and about 5% in a seven‐dimensional space. Speech synthesized from the vector quantized principal components using nine bits per frame was somewhat inferior in quality to speech synthesized with full precision principal components, but quite intelligible. [Work supported by NSF.]