Speech Analysis-Synthesis and Recognition System.

Abstract

In order to examine the effectiveness of parameters in speech recognition, it is preferable to reproduce speech sound from the extracted parameters by a so-called speech analysis-synthesis system. From this point of view, we have constructed a composite speech research system which consists of a speech analysis-synthesis part and a speech recognition part. Both of them are based on the zero-crossing analysis method. Experiments were conducted with 1900 words produced by five adult males. The average information compression ratio of the analysis-synthesis part was about 1/10 compared with 7 bit PCM coding. Intelligibility of the synthesized speech was ascertained to be rather good. In the recognition part, every segmented part of speech is recognized as either one of the 5 vowels and 8 consonant groups. About 95.6% of the vowels and 69.0% of the voiced consonants were recognized correctly by personally adjusted discriminats. Up to 86.9% of the unvoiced consonants could be identified correctly. Besides, extended recognition experiments were conducted and in these experiments common discriminant functions were applied to all the subjects and other ten male students. The results obtained made clear the efficiency and the merit of the recognition scheme.