Decision Functions for Voiced-Unvoiced-Silence Detection
Carol C. Lochbaum, Edward E. David, M. V. Mathews · The Journal of the Acoustical Society of America · 1961
At the June, 1960, Acoustical Society of America meeting, we described a voiced-unvoiced-silence detection procedure. It incorporated two independent threshold decision rules, one operating on the zero moment of the speech spectrum for the spoken-silent choice, the other operating on the first moment for the voiced-unvoiced choice. A generalization of this procedure involves a decision which depends jointly on the two moments. This paper compares the performance of this process with the simpler one. Value functions placing specified weights on the nine possible outcomes of this 3×3 detection situation were defined, and decision rules maximizing the values were computed. Maximum likelihood, maximum likelihood weighted by prior probabilities, and rules tailored to vocoder applications were studied. Probabilities of the six possible confusions and the three correct answers were computed for all rules. Data for the evaluation consisted of twenty 3-sec utterances by two men and two women. Voiced, unvoiced, and silent intervals in these speech samples were established visually from spectrograms. Performance depends, of course, on the value functions, but at most a reduction of 13 in the over-all error rate over the simple rule was achieved.