HMM-based noisy-speech pitch contour estimation

A robust pitch contour estimation algorithm that makes use of hidden Markov models (HMMs) is proposed for speech signals corrupted by various kinds of noise at low SNR. The HMM contains three states. For the output p.d.f., multivariate continuous Gaussian distributions are chosen, having pitch and its derivatives as the random variables. Pitch estimation via two steps is suggested so that information among inter- and intra-analysis frames can be fully used. Conventional algorithms are thus modified to provide a group of weighted pitch candidates. For all possible candidate sequences with given a priori probabilities, the HMM-based algorithm calculates the output probabilities and selects the best contour from the candidates in the maximum-likelihood (ML) sense. To solve the problems of excessively using memory space and missing wanted candidates, beam search, pruning and candidate prediction are used. Advantages of HMM estimation over the conventional smoothing and dynamic programming (DP) are also discussed. The preliminary simulations showed its robustness