Variable dimension vector quantization of linear predictive coefficients of speech

摘要

We introduce a method for locally optimal variable-to-variable length source coding with distortion, and apply it to coding the linear predictive coefficients of speech. The method is similar to entropy-constrained vector quantization, but it uses a dynamic programming algorithm to encode. The method automatically discovers variable-length source structure, in this case the acoustic-phonetic structure of speech. Using this structure, it is possible to compress the linear predictive coefficients of speech to one-third the rate of entropy-constrained vector quantization of speech, with no increase in spectral distortion. Auditory tests reveal that using this method, the spectral component of speech can be coded naturally and intelligibly to as low as 50 bits per second.