With the fast development of embedded technology, embedded devices are getting closer and closer to human being’s daily life. These devices usually appear in the form of information terminals, which not only integrate multiple functions together, such as computing, communication and sensor, but also could be conveniently connected to conventional devices and equipments. Therefore, to satisfy users’ demand for much easier and more natural access to embedded devices, it is necessary to integrate text-to-speech function into human-machine interface. However, as the research for speech science starts late in China, currently the embedded mandarin speech synthesis technology is far from advanced yet and there are few related mature products. Another reason for the limitation of embedded TTS application is that compared with western language, the prosody model for Mandarin TTS is usually more complicated and needs more text and voice resources to achieve a highly natural result, thus making the system sensitive to its environment. The goal of this paper is to implement an applicable embedded mandarin TTS system which has a commercialization potential. Based on the original syllable-based large scale speech corpus, it innovatively introduces non-uniform unit as the basic system unit, and proposes a data mining based voice library tailoring strategy. It first analyzes the original voice library based on statistics of acoustic features, and then utilizes data mining techniques to do appropriate pruning job, trying to keep as much prosodic and acoustic coverage as possible in the tailored target voice library. The method achieves a big compression ratio while the output speech based on target library remains natural and intelligible. With this method, several small target speech databases are built for different embedded environments. The TTS system based on those databases consumes little resources, but outputs clear and natural speech. Besides, it could be easily adapted to various embedded environments.
修改评论