The higher-order statistics of speech signals

Most techniques used in speech processing are based on the second-order properties of the speech signal, described by the power spectrum, spectrogram, and autocorrelation function. However, there is generally more information in a speech signal than second-order analysis can provide, and the study of the signal´s higher-order statistics (HOS) can reveal useful information about the speech production process. By using HOS techniques non-linearities in the speech production process can be identified, and these may be used to modify the linear speech model to produce more realistic speech. The third-order statistics, described by the triple correlation and bispectrum, can give information about the signal skewness, and can identify the presence of quadratic phase coupling in the signal. The authors present preliminary results from a pitch-synchronous bispectral analysis of sustained vowels, and show that the HOS properties are of significant size to warrant further investigation