Pitch Synchronous Spectral Analysis Scheme for Voiced Speech

Abstract

This paper addresses the problem of spectrum analysis of harmonic signals which are periodic or at least quasi-periodic, such as human voice. The Fourier coefficients uniquely represent such signals and thus fully characterize them. An efficient and accurate pitch synchronized spectral analysis scheme to obtain the Fourier coefficients of a harmonic signal, sampled at an arbitrary rate above the Nyquist critical rate, is outlined. Since the pitch is not known a priori, it is derived from the sampled signal prior to the spectral analysis. The rationale behind the proposed scheme is based on an interpolation of the signal with an upsampling rate that is synchronized with the pitch period of the signal. If the resulting upsampled sequence is periodic, its Fourier coefficients can be easily computed using a Discrete Fourier Transform analysis of a single pitch period. However, it is shown that the resulting upsampled sequence is aperiodic, but nevertheless may be decomposed into a periodic signal corrupted by a small aperiodic high-frequency noise. The fact that this noise is correlated with the signal is employed to obtain a closed-form solution for the desired Fourier coefficients from the noisy values, using the computationally superior FFT algorithm. The accuracy of the proposed scheme is demonstrated for synthetic speech where the spectrum is known a priori. The results obtained for real speech signals show better consistency across adjacent frames as compared to conventional methods. © 1989 IEEE