HMM-based speech synthesis with a flexible Mandarin stress adaptation model

Expressive speech synthesis has recently received much attention. Stress is one key issue which may improve the expressiveness of the synthetic speech. However, rare work was done in Mandarin stress prediction and expression. This paper presents a HMM-based expressive speech synthesis system which supports Mandarin stress synthesis. Mandarin stress was automatically predicted with textual features only using a Maximum Entropy Model. The linear adaptation model was extracted from a large corpus by analyzing their stress related acoustic features. The advantage of the proposed model is it can be easily modified to build a system with another speaking style or emotion. Experiments show that the proposed stress adaptation system can convey stress effectively and generate high expressive speech. The overall performance of the synthetic speech is also improved.