The experimental results of applying a simple method for automatic segmentation of speech are presented. The performance of the segmentor is that it provides boundary indications for elementary speech sounds with more than 75% reliability, measured over the utterances of four male speakers. The criterion of performance is that the segmentor should automatically provide in real time the boundaries between elementary speech sounds (i.e., phonemes) that are the same as those obtained in the analysis of a sonogram of the utterance by a trained human observer. The main source of error in the present segmentor is not incorrect boundaries, but extra boundaries inserted between correct boundaries. This fact makes the effective accuracy of the segmentor, for the purpose of automatic recognition of phonemes, better than its 75%–80% absolute accuracy.

This content is only available via PDF.
You do not currently have access to this content.