Application of Bayesian Autoregressive Detector for Speech Segmentation

This contribution addresses the basic properties of the Bayesian Autoregressive Changepoint Detector (BACD) and its use for speech segmentation. The principle of the BACD consists in the identification of changes in both the voice excitation and vocal tract parameters. Thus different piecewise autoregressive (AR) models with changes in the order and coefficients are used to describe speech units. By searching for a maximum of a posteriori density (MAP) given by the BACD the boundary of each speech unit can be found. The use of the BACD for text-to-speech analysis-synthesis and the improvement of labelling corpora are considered as the primary applications of this approach.