On the use of a perceptual distance with speech vector quantization
René Carré, Christian Lacoste, D. Tuffelli · The Journal of the Acoustical Society of America · 1984
A vector quantization using perceptual data has been studied. The frequency and bandwidth scales of the poles computed by LPC methods were quantized according to perceptual difference limen. The small sensitivity of the hearing system to the formant amplitude was thus taken into account. A spectrum dictionary was then obtained. One spectrum is a set of quantized poles and the inter-spectrum distance is either 1 or 0. Good results are expected in the case of multi-speaker applications: low bit rate transmission and speech recognition. The method is well adapted for controlling all the steps of the processing and thus for improving the obtained results. For instance, the evolution of the dictionary can be easily studied with an increasing number of speakers and the statistical importance of each spectrum can be measured. First synthesis are quite good and comparable with those obtained by classical vector quantization. A phonetic labelling of the signal will lead to a dictionary spectrum labelling and a pointer giving the number of each successor permits to study the coarticulation effects at the level of the spectra. [Work supported by DRET.]