An intelligibility evaluation of several linear prediction vocoder modifications

Abstract

Two linear prediction vocoder modifications and their intelligibility evaluation are presented. Lower order coefficient representation and faster frame rate analysis during unvoiced speech are proposed for intelligibility improvement without increased transmission rates. Diagnostic Rhyme Test (DRT) comparisons between (a) the fixed filter order, fixed analysis frame rate vocoder, (b) various modified vocoders, and (c) the input speech, are implemented using two speakers, seven listeners, and a selected set of word pairs. Reducing the filter order from 10 to 4 for unvoiced speech frames analyzed at a rate of 44.4 frames/sec produces no significant decrease in the scores for all of the distinctive features tested. Updating only the unvoiced analysis frames at twice the frame rate (88.8 frames/sec) leads to a significant improvement in the scores for the sustention feature. Interpretations of the DRT results, and considerations for implementing the test within ordinary laboratory facilities are discussed.