Syllable Detection in Continuous Speech

Abstract

A computer program for syllable detection was developed for possible application in future speaker independent speech recognition systems. Usage of both peak-to-peak (P-P) amplitude and rms intensity measurements was considered, but it was observed that P-P amplitude measurements alone provided good results, and that almost none of the resulting syllable detection errors could be eliminated through the additional consideration of rms intensity information. Syllable detection was based upon the recognition of the variations in amplitude which frequently accompany the onset and offset of both syllable nuclei and associated nonvowel sonorants. Changes in P-P amplitude and the maximum amplitude occurring within each change were therefore the primary measurements considered in the recognition process. Thus far, an averaged syllable detection accuracy of 86% has been obtained for nine talkers speaking a total of 3.5 min of continuous speech, with individual syllable detection results ranging from 74% to 96%. Misclassification of nonvowel sonorants comprised about 80% of the syllable detection errors, and it is consequently in this area of the program's performance that improvements are currently being considered.