Research On Emotional Speech Synthesis And System Building

Speech synthesis is one of the most important parts in human-machine communication. Currently, speech synthesis has been developed to an extent with considerable intelligibility. However, speech as a way of man-to-man message conveyance can hardly work well without the involvement of emotions. The proposal of emotional speech synthesis has brought in a higher demand on conventional speech synthesis.The research is under a joint project with Waseda University, aiming at building a robot which can make emotional speech in Chinese, English and Japanese. Here, we mainly focus on the mandarin emotional speech synthesis.This thesis first introduces the development of speech synthesis, the status of this area and challenges it faces. Then, the thesis recaps and compares existing fundamental synthesis methods. Given the experiment availabilities, the concatenative method is chosen for our basic speech synthesis. In the corpus building stage, we fully employed the features of mandarin and took into consideration the position information of single words in phrases and phrases in a whole sentence.Based on some prevalent emotion theories, experiments on emotional mandarin speech samples are carried out and the emotion features and acoustic parameters are obtained after analysis. Besides, detailed extraction methods for each parameter are described and PSOLA algorithm and its use in our project are elaborated on. Finally, a completed mandarin emotional speech synthesis system is given, which is divided into such two stages as neutral speech synthesis and emotional speech generation. The system is advantageous in its small size and appropriate emotion presentation.