Discrimination of five voice qualities and prediction to perceptual ratings.

Abstract

A series of acoustic measures based on the long-term average spectrum were developed to discriminate five voice qualities. These were breathy voice, creaky voice, nasal voice, tense voice and whispery voice. A preliminary study first tested the effectiveness of a series of 'equal-Hertz' intervals in terms of their ability to discriminate the five voice types. The best of these was then compared to Bark and Mel intervals, using discriminant analysis. The most effective discriminators were shown to be the 1.5-Bark and 200-Mel intervals. The significant discriminators in all the discriminant analyses were then used to predict expert perceptual ratings of the voice qualities. All measures developed successfully predicted to all the voice types except nasal voice.