Variable-rate speech coding: Replacing unvoiced excitations by linear prediction residues of different phonemes
Wolfram Ehnert, Ulrich Heute · 1997
In order to reduce the bit rate of speech transmission while maintaining the speech quality we are developing a vocoder which uses different methods for coding voiced and unvoiced frames. Within this framework we present an idea for expressing fricative and plosive phonemes with only 20 bits per frame (tD 20ms). We show that they can be represented by LP(Linear Prediction) coefficients and a residual signal where this residue is always taken from a fixed phoneme of a test speaker known at the receiver station of the coding system (see figure 1). Algorithms ensuring smooth transitions to other speech-frame categories are also described below. Using this technique the transmission rate of unvoiced frames can be considerably reduced (down to 1 kbit/s) getting better listening results than using CELP (Code Excited Linear Prediction) variants at 4 kbit/s instead. The resulting ‘Multi-Class Vocoder’ (voiced frames are coded by Harmonic Coding at 4 kbit/s) has a variable rate of less than 3 kbit/s on the average.