Maximum likelihood pitch estimation

The problem of determining the pitch period of voiced speech is formulated as a maximum likelihood (ML) estimation problem for an unknown periodic signal in white Gaussian noise of unknown intensity. Modifications of the ML estimator for speech include removal of bias, derivation of a measure of confidence, and prefiltering for non-white noise. The pitch estimation procedure is evaluated in the frequency domain and related to comb filtering. An alternate derivation of the ML estimator, related to the procedure proposed by Lacoss for maximum likelihood spectral estimation is presented. The problem formulated is to find a filter which passes a periodic signal of period P without distortion while simultaneously suppressing, in an optimum manner, a noise signal having a known spectrum.