Vector quantization of LPC parameters based on dynamical features of hearing
S. Ono, Takashi Araseki · The Journal of the Acoustical Society of America · 1987
New code-word search criteria for vector quantization of LPC parameters are proposed. Various vector quantization schemes for LPC parameters have been proposed previously. In these schemes, an input speech waveform is partitioned into successive short time frames and code-word search is carried out under the criteria based on instantaneous speech features. Characteristics of quantization noise vary widely from frame to frame, independent of the dynamic characteristics of the input speech. At very low bit rates, where it is difficult to keep the quantization distortion at a low level, variation in the quantization noise significantly degrades the quality of the reconstructred speech. “Temporal masking” experiments indicate that there is a postive correlation beteen the masking value and the speech spectral transition rate. These experiments also suggest that spectral transition rate distortion is emphatically perceived. In this paper, new criteria, based on these temporal masking properties, are proposed. They would minimize not only the instantaneous spectral distortion, but also the spectral transition rate distortion. Using these criteria, vector quantization of the LPC parameters is developed. Subjective experiments confirm that these criteria are superior to those based only on instantaneous speech features.