Characteristics of pitch extraction by ACLOS (AutoCorrelation of LOg Spectrum)

Abstract

A new pitch extraction method named ACLOS (AutoCorrelation of LOg Spectrum) [Kunieda et al., ICASSP-96, pp. 232–235 (1996)] is proposed. In this method, the fundamental frequency is estimated from the maximuum peak of the autocorrelation function of the log spectrum. ACLOS gives more robust and reasonable pitch information than other methods give. Quantitative evaluation of pitch extraction, however, has not been carried out. This paper presents a comparative performance of pitch extraction by ACLOS, the autocorrelation method, and the cepstrum method based on the assessment method employed by Rabiner [Rabiner et al., IEEE Trans. Acoust. Speech Signal Process. ASSP-24, 399–418 (1976)]. In order to increase the robustness for noisy speech, ACLOS is modified. Clipping of the log spectrum before calculating the autocorrelation function is introduced, because noise remarkably affects the valley of the log spectrum in low SNR. Experimental results reveal that ACLOS can reduce the gross pitch error even if the speech signal is degraded by noise. Conclusively, it is indicated that ACLOS will be a powerful tool for pitch extraction of noisy speech.

Frequency and Voice: Perspectives in the Time Domain

“…Notable variations and speed improvements of the autocorrelation method have been reported. 10,24,25 As with ZC and PD methods, associating a sinusoidal model with the cycle lengths obtained by AC provides for the signal to be described in terms of frequency. For some signals with high variability in amplitude, covariance may be more appropriate than correlation in determining cycle length.…”

Using a new Discretization of the Fourier Transform to Discriminate Voiced From Unvoiced Speech

“…Now the autocorrelation can be computed and the periodicity peaks wont be confused with the peaks due to vocal tract oscillations. The full algorithm for determining the nature of the signal (voiced/unvoiced) and its pitch is according to [8], [4] 2. Find the largest peak of the first and the last 10 msec of the segment.…”