Comparison of pitch detection by cepstrum and spectral comb analysis

Abstract

Most reliable methods of detecting pitch in the speech signal are based on the assumed periodicity found in the voiced speech spectrum (cf. the cepstrum method). Due to recent hardware developments, this approach is becoming more attractive since the prerequired harmonic analysis can now be easily performed in real time. The spectral comb correlation, a new method based on spectral analysis, was recently introduced (5). Rather than performing a second spectral analysis on the log power spectrum, this approach relies on the correlation between the power spectrum and a spectral comb with of decreasing amplitude and variable teeth intervals. The maximum of this correlation function is obtained when the width of the teeth intervals corresponds to the periodicity of the spectrum. The spectral comb method is compared to the cepstrum on both theoretical and experimental levels. It is shown in particular that the comb method is computationally more efficient, gives a better fundamental frequency resolution, and can use smaller duration windows for the initial spectral analysis.