Nebula: F0 Estimation and Voicing Detection by Modeling the Statistical Properties of Feature Extractors

A F0 and voicing status estimation algorithm for speech analysis/synthesis is proposed. Instead of directly modeling speech signals, the proposed algorithm models the behavior of feature extractors under additive noise using a bank of Gaussian mixture models, trained on artificial data generated from Monte-Carlo simulations. The conditional distributions of F0 predicted by the GMMs are combined to generate a likelihood map, which is then smoothed by a Viterbi search to give the final F0 trajectory. The voicing decision is obtained based on the peak F0 likelihood. The proposed method achieves an average F0 gross error of 0.30% on CSTR and CMU Arctic speech datasets.