An information theoretic perspective on the speech spectrum process

Abstract

In this paper, an information theoretic study of properties of the speech spectrum process is performed. Various techniques to model the probability density function are applied to the spectrum source to compute rate-distortion functions. We estimate the difference in the required rate to achieve a given distortion for three different scenarios: interframe gain exploitation, low-pass filtering of LPC vectors and increased speech signal bandwidth. We obtain fairly consistent results for the different methods of calculating rate-distortion functions. The results show that for close to transparent LPC quantization we gain 4-6 bits per frame by exploiting first order interframe correlation. The new idea of using low-pass filtered LPC vectors has shown to decrease the coding cost with 1-3 bits per frame, depending on the cutoff frequency.