Cross-Validation and Minimum Generation Error based Decision Tree Pruning for HMM-based Speech Synthesis

This paper presents a decision tree pruning method for the model clustering of HMM-based parametric speech synthesis by cross-validation (CV) under the minimum generation error (MGE) criterion. Decision-tree-based model clustering is an important component in the training process of an HMM based speech synthesis system. Conventionally, the maximum likelihood (ML) criterion is employed to choose the optimal contextual question from the question set for each tree node split and the minimum description length (MDL) principle is introduced as the stopping criterion to prevent building overly large tree models. Nevertheless, the MDL criterion is derived based on an asymptotic assumption and is problematic in theory when the size of the training data set is not large enough. Besides, inconsistency exists between the MDL criterion and the aim of speech synthesis. Therefore, a minimum cross generation error (MCGE) based decision tree pruning method for HMM-based speech synthesis is proposed in this paper. The initial decision tree is trained by MDL clustering with a factor estimated using the MCGE criterion by cross-validation. Then the decision tree size is tuned by backing-off or splitting each leaf node iteratively to minimize a cross generation error, which is defined to present the sum of generation errors calculated for all training sentences using cross-validation. Objective and subjective evaluation results show that the proposed method outperforms the conventional MDL-based model clustering method significantly.