Toward improving speech intelligibility

Abstract

One difference observed between conversational and clear speech appears to be the presence of greater amounts of vowel reduction and target formant frequency “undershoot” in conversational speech. Thus vowel formant frequencies in clear speech more closely reach targetlike values, and transitions out of semivowels are steeper in slope [Picheny and Durlach (1979)]. To begin examining the contribution of such effects to the overall improvement in intelligibility of clear speech, an algorithm has been developed which, approximately, inverts the effects of reduction and undershoot in conversational speech. The algorithm is applied to the parameters used by an LPC vocoder program so that, in three steps, synthesis parameters are obtained via LPC analysis, changes in the time course of each synthesis parameter are magnified, and a synthetic utterance is generated which contains enhanced time‐changing information. Thus the algorithm operates on spectral change in the original version of an utterance, and magnifi...