Comparing qmt1 and HMMS for the synthesis of American English prosody

Abstract

Three models are compared for the duration and pitch contour of American English in a speech synthesis framework. These models combine duration prediction by Quantification Method Type 1 (QMT1), a Codebook-based method for the F0 contour and a Hidden Markov Model-based method for both durations and F0. Subjective listening tests show that the HMMs are preferred over the Codebook for the F0 contour, but that their duration modelling performances are not significantly different from those of QMT1 in the tested setup. An analysis of naive freeform listener comments supports this fact, and suggests that such comments can give useful hints regarding the performance of each system.