A semiautomatic pitch detector (SAPD)

The purpose of this paper is to describe a technique for semiautomatically determining the pitch contour of an utterance. The method is significantly more sophisticated than the standard technique of hand tracking of pitch periods from a waveform display of the utterance and leads to a fairly robust measurement of the pitch period. This technique utilizes a simultaneous display (on a 10 ms section-by-section basis) of the low-pass filtered waveform, the autocorrelation of a 400- point segment of the low-pass filtered waveform, and the cepstrum of the same 400-point segment of the wideband recording. For each of the separate displays (i.e., waveform, autocorrelation, and cepstrum) an independent estimate of the pitch period is made on an interactive basis with the computer, and the final pitch period decision is made by the user based on results of each of the measurements. The technique has been tested on a large number of utterances spoken by a variety of speakers with very good results. Formal tests of the method were made in which four people were asked to use the method on three different utterances, and their results were then compared. During voiced regions, the standard deviation in the value of the pitch period was about 0.5 samples across the four people. The standard deviation of the location of the time at which voiced regions became unvoiced, and vice versa was on the order of half a section duration, or 5 ms. The major limitation of the proposed method is that it requires about 30 min to analyze 1 s of speech. However, the increased accuracy and robustness of the results indicate that the tradeoff of time for accuracy is a good one for many applications.