Comparative performance study of several pitch detection algorithms

Abstract

A comparative performance study of seven pitch detection algorithms was conducted. A speech data base, consisting of eight utterances spoken by three males, three females, and one child was constructed. Telephone, close talking microphone, and wideband recordings were made of each of the utterances. For each of the utterances in the data base; a "standard" pitch contour was semiautomatically measured using a highly sophisticated interactive pitch detection program. The "standard" pitch contour was then compared with the pitch contour that was obtained from each of the seven programmed pitch detectors. The algorithms used in this study were 1) a center clipping, infinite-peak clipping, modified autocorrelation method

A comparative performance study of several pitch detection algorithms

“…Some of the raw analysis results for speaker Dl are presented in the M.S. thesis of Cheng [18]. Finally, because of the large number of factors involved in the analysis, each of the error measurements was averaged over the utterances of each speaker.…”

Real-time digital hardware pitch detector

“…The hardware pitch detector described in Section III has been built and interfaced to the NOVA computer facility of the Acoustics Research Department. An extensive performance evaluation was made of the capabilities of this and several other pitch-detection algorithms using software simulations [9]. To test the performance of these algorithms, a speech data base, consisting of eight utterances spoken by 3 males, 3 females, and 1 child was constructed.…”

“…When the pitch period exceeds 150 samples, the analysis frame size is not large enough to hold two full periods of speech; thereby increasing the chance of a gross error in locating the correct pitch period. In the study of [9], the two low-pitch speakers used in the study both showed a large number of gross errors in the pitch period (for both the telephone and microphone recordings). All other speakers had only occassional gross errors in the pitch period (i.e., one gross error/s on average).…”

“…All other speakers had only occassional gross errors in the pitch period (i.e., one gross error/s on average). A nonlinear smoothing algorithm [11] was used in the study of [9] to isolate and correct these gross errors, as well as isolated errors in the voiced-unvoiced decision. After processing by the nonlinear smoothing algorithm, essentially all gross pitch-period errors were corrected except for the case of the low-pitch male speaker where some of the errors occurred in clusters, and therefore were essentially not correctable by a median-type smoother.…”

“…The second category in the performance evaluation of [9] was the accuracy in voiced-unvoiced boundary location. For the autocorrelation pitch detector, the average error in locating the voiced-unvoiced boundary was on the order of 5 ms (half the average frame rate of 100 frames/s or 10 ms/ frame), and the standard deviation of the error was on the order of 10 ms across all speakers and recording conditions.…”