| 摘要 |
The pitch of human speech segments is analyzed using at least two different pitch detection algorithms, a respective plurality of most likely values of pitch is selected by each of those algorithms, and these values and their respective quality figures are analyzed statistically to determine the most likely pitch. One algorithm operates in the frequency domain, by analyzing an amplitude spectrum, and the other algorithm operates in the time domain using an autocorrelation function. Significant peak positions of the amplitude spectrum or autocorrelation function are evaluated in respective harmonic sieves, to provide respective quality figures indicating the degree to which peak frequency or period periods of the spectrum or autocorrelation function output match the apertures of the harmonic sieve. A predetermined number of values of pitch, and of period, are selected having the highest quality figures. After conversion of the values for period into values of pitch, these values with their associated quality figures are analyzed statistically to form an estimation of the most likely pitch.
|
