Cross validation and Minimum Generation Error for improved model clustering in HMM-based TTS

Abstract

In HMM-based speech synthesis, context-dependent hidden Markov model (HMM) is widely used for its capability to synthesize highly intelligible and fairly smooth speech. However, to train HMMs of all possible contexts well is difficult, or even impossible, due to the intrinsic, insufficient training data coverage problem. As a result, thus trained models may over fit and their capability in predicting any unseen context in test is highly restricted. Recently cross-validation (CV) has been explored and applied to the decision tree-based clustering with the Maximum-Likelihood (ML) criterion and showed improved robustness in TTS synthesis. In this paper we generalize CV to decision tree clustering but with a different, Minimum Generation Error (MGE), criterion. Experimental results show that the generalization to MGE results in better TTS synthesis performance than that of the baseline systems.