Short-Time Spectrum and “Cepstrum” Techniques for Vocal-Pitch Detection

Abstract

A spectrum analyzer based on a definition of short-time power spectra has been designed and simulated on a digital computer. The analyzer is primarily intended for use in speech analysis. It has been designed to operate in real time, and to produce high-resolution spectra without utilizing either heterodyning methods or bandpass filter banks. The logarithm of each consecutive amplitude spectrum thus obtained can be used as the input to a second similar spectrum analyzer. The output of this analyzer is then the “cepstrum” or power spectrum of the logarithm spectrum. The cepstrum of a speech signal has a peak corresponding to the fundamental period for voiced speech but no peak for unvoiced speech. Thus, a cepstrum analyzer can function both as a pitch and as a voiced-unvoiced detector. Cepstral pitch detection has the important advantages that it is insensitive to phase distortion, and is also resistant to additive noise and amplitude distortion of the speech signal. The method does not require the presence of the fundamental frequency in the speech signal, and will give several separate cepstral peaks if several different pitch periods are present. Cepstral techniques appear to be even more reliable and efficient than visual methods for pitch detection. The short-time spectrum and cepstrum analyzers described in this paper were simulated by a sampled-data system on an IBM-7090 digital computer. The simulation was programmed with the assistance of a special block-diagram compiler.