System for Automatic Formant Analysis of Voiced Speech

Abstract

A system for automatically estimating the lowest three formants and the pitch period of voiced speech is presented. The system is based on a digital computation of the cepstrum (defined as the inverse transform of the log magnitude of the z-transform). The pitch period estimate and smoothed log magnitude are obtained from the cepstrum. Formants are estimated from the smoothed spectral envelope using constraints on formant frequency ranges and relative levels of spectral peaks at the formant frequencies. These constraints allow the detection of cases where two formants are too close together in frequency to be resolved in the initial spectral envelope. In these cases, a new spectral analysis algorithm (the chirp z-transform algorithm) allows the efficient computation of a narrow-band spectrum in which the formant resolution is enhanced. Formant and pitch period data obtained by the analysis system are used to control a digital formant synthesizer. Results, in the form of spectrograms, are presented to illustrate the performance of the system.