Acoustic correlates of intervocalic stop confusions

Abstract

As part of the study of signal processing algorithms that automatically enhance speech intelligibility, the acoustic factors underlying place confusions in intervocalic stop consonants have been examined. The speech database in these studies consists of nine repetitions of each of eighteen nonsense utterance frames from five talkers. The utterance frames were of two forms: “say a CVwuh” and “say a wVCuh” with C = {/b/,/d/,/g/} and V = {/i/,/a/,/u/}. In producing the repetitions of each frame, talkers were under instructions to vary their speaking style from enunciating very clearly to speaking in a casual, relaxed fashion. Confusion matrices and percentage correct consonant identification have been obtained for these utterances from both normal‐hearing listeners (using broadband noise as a masker) and from hearing‐impaired listeners. Acoustic measures related to place of articulation cues and general spectral/temporal contrast have been obtained from each utterance. The relations between these acoustic measures and consonant confusability will be described and implications for algorithms intended to enhance speech intelligibility will be discussed. [Work supported by NINCDS.]