基于HMM的英语焦点语音合成

Emphasis is an important form of expressiveness in speech. Hidden Markov model (HMM) based synthesis has shown great flexibility in generating expressive speech. However it does not work well when the training data is lacking as in the case of emphatic speech where a sentence usually contains only few emphatic words. To address the issue, this paper proposes a modified two-pass decision tree with a compensation model for emphatic speech synthesis. The English syllables are first categorized into 6 classes based on their locations and distances in relation with the nearest emphatic word and its primary stressed syllable(s). The English phones are then grouped into 9 types according to their pronunciation method. The decision tree is then constructed using non-emphasis related context questions in the first pass, and is then extended with emphasis questions in the second pass. The compensation model is used to modify the parameters predicted by the two-pass decision tree based on phone types rather than individual phones to alleviate the sparse data problem. Experiments show that the proposed method can generate emphatic speech with high quality for both naturalness and emphasis.