Automatic Estimation of Formant Frequencies for Voiced Speech

Abstract

A system has been developed for automatically estimating the lowest three formants and the pitch period for voiced speech. The cepstrum is used to obtain an estimate of the pitch period and the spectral envelope. The algorithm for estimating the formants from the spectral envelope incorporates constraints on formant-frequency ranges and relative spectral levels. When two formants are close together and are found to be unresolved in the initial spectral envelope, a new spectral analysis algorithm is used to resolve the two formants. The formant and pitch-period data are used to control a digital formant synthesizer. Examples of the synthetic speech produced from the estimated formant and pitch data are presented.