Predicting consonant confusions from acoustic analysis

Acoustic measurements of nonsense syllables in quiet and in noise were used to predict the pattern of consonant confusions made under those conditions. Eleven sets of nonsense syllables were presented to normal-hearing subjects in quiet and at a +5 dB speech-to-noise ratio, at five speech levels. A set of acoustic characteristics of the speech stimuli were chosen for analysis and measured using digital processing techniques. Results of the recognition task revealed significant effects of consonant voicing, position and vowel context on syllable recognition. The performance-intensity function of the quiet condition rises more steeply than the function obtained in noise. The effect of noise on consonant recognition is dependent upon the manner in which the consonant is produced, and the location of maximum constriction. Differences in the absolute values of the acoustic parameters of syllable pairs were used to predict their percentage of confusion. A set of acoustic variables was isolated which was found to be the best predictor of confusion percentages. Although the sets of acoustic variables were different for various syllable types and test conditions, three variables (consonant energy, consonant spectral peaks, consonant-to-noise ratio) were used in a majority of the predictions.