Cross-spectral methods with applications to speech processing
Douglas J. Nelson · The Journal of the Acoustical Society of America · 1999
Cross-spectral methods were first presented in the mid-1980’s as a method for accurately estimating stable parameters such as modem baud rate and carrier frequencies. The stability of these signal parameters makes it possible to integrate large amounts of data to accurately estimate parameters even under degraded conditions. Since biological signals, such as speech, are not stationary, classical analysis methods, including normal cross-spectral methods, are poorly suited to the problem. Presented here are methods which take advantage of the structure of speech and the phase properties of the Fourier transform. They are based on the cross-spectral methods of the 80’s, but have the advantage that these newer methods provide good accuracy and resolution for nonstationary signals such as speech. In addition, they provide a simple method for taking advantage of signal structure, such as the harmonic properties of speech, which results from the quasi-periodic pitch excitation. Specific problems addressed are accurate pitch and formant estimation, and problems such as blind recovery of carrier frequencies for single side band AM radio transmissions.