JNDS IN THE LPC POLES OF SPEECH AND THEIR APPLICATION TO QUANTIZATION OF THE LPC FILTER

The poles of a linear prediction coder (LPC) are directly related to speech formants; thus, quantizing the poles in an LPC vocoder is the most direct method for incorporating knowledge on formant perception in the vocoder design. For the purpose of guiding a pole quantization scheme, a psychophysical experiment was performed to measure just noticeable differences (JND's) in the frequency and radius of the poles. The results can be associated with formant JND's, extending presently existing data. The frequency JND's, measured up to a formant frequency of 4 kHz, are quantified as distributions with means that are increasing functions of formant frequency and band-width. An example of a pole quantization scheme, based on the JND data, is presented and found to be significantly superior to common scalar quantization methods of the LPC-PARCOR coefficients. The pole quantization scheme is also found to be almost comparable, both in quality and bit consumption, to vector quantization.