An analysis-by-synthesis approach for automatic time segmentation of speech signals

This paper presents an analysis-by-synthesis speech segmentation algorithm. The approach employs a variable-frame-length LPC speech synthesizer, and the segmentation process is performed by constantly monitoring the synthesis output. The resulting segments, in general, do not correspond to elemental units of speech such as phonemes, but rather they represent a minimum number of acoustic units necessary for a high-quality reproduction of the particular utterance. On a limited speech data base, synthetic continuous speech with an average rate of 20 frames per second and with quality comparable to about 333 frames per second fixed rate LPC speech has been achieved.