Robust pitch determination of noisy speech.
J.A. Rodriguez-Fonollosa, Asunción Moreno · The Journal of the Acoustical Society of America · 1991
Most of the signal analyses made up to date have been based on the autocorrelation function or power spectrum. It is well known that these second-order statistics completely characterize a Gaussian process. However, in many applications where non-Gaussian processes or nonlinearities are present, analysis based on the autocorrelation (and hence, the power spectrum) fails to provide all the useful information about the process [C. L. Nikias and M. R. Raghuver, Proc. IEEE 75, 869–891 (1987)]. In this paper, it is shown that higher-order statistics are very useful in the study and characterization of the speech for two reasons: They can extract useful information about the statistics of voiced frames, and they can separate speech from noise. Taking advantage of these properties, a new pitch determination algorithm based on third-order statistics has been developed. Third-order statistics are quite insensitive to most noises (Gaussian, sinusoidal, car noise,...) because these noises have a symmetric probability density function, and therefore, their third-order cumulants are zero. The algorithm has been tested in noise-corrupted speech, at different levels of signal-to-noise ratio, and with different kinds of noise. The results show that this new algorithm gives a much better estimation of the pitch than the conventional autocorrelation method in all the cases. For example, the cumulant-based detector reduces the gross pitch error rate from 16.8% to 5.4% for speech sentences corrupted with car noise and a signal to noise ratio of 5 dB.