5th International Conference on Spoken Language Processing

Sydney, Australia
November 30 - December 4, 1998

Towards A Chinese Text-To-Speech System With Higher Naturalness

Ren-Hua Wang, Qinfeng Liu, Yongsheng Teng, Deyu Xia

University of Science & Technology of China, China

This paper presents our research efforts on Chinese text-to-speech towards higher naturalness. The main results can be summarized as follows: 1. In the proposed TTS system the syllable-sized units were cut out from the real recorded speech, the synthetic speech was generated by concatenating these units back together. 2. The integration of units synthesized by rules with natural units was tested. A LMA filter based synthesizer was developed successfully to test and generate those units, which were difficult to be collected from the speech corpus. 3. A new efficient Chinese character coding scheme - "Yin Xu Code"(YX Code) has been developed to assist the GB Code . With the YX Code a new lexicon structure was designed up. The new dictionary system not only supplies with the pronunciation information, but also is much helpful for the words-segmentation. Based on above results, a Chinese text-to-speech system named as "KD-863" has been developed. The system converts any Chinese written text to speech in real time with high naturalness. In the national assessment of Chinese TTS systems held at the end of March 1998 in Beijing, the system achieved a first of the naturalness MOS (Mean Opinion Score).

Full Paper

Bibliographic reference.  Wang, Ren-Hua / Liu, Qinfeng / Teng, Yongsheng / Xia, Deyu (1998): "Towards a Chinese text-to-speech system with higher naturalness", In ICSLP-1998, paper 0172.