On unit analysis for Cantonese corpus-based TTS

This paper reports a study of unit analysis for concatenative TTS, which usually has an inventory of hundreds of thousand of voice units. It is known that the quality of synthesis units is especially critical to the quality of resulting corpus-based TTS system. This research focuses on the analysis of a Chinese Cantonese unit inventory, which has been built earlier for open vocabulary Chinese Cantonese TTS tasks. The analysis results show that the exercise helps identify the sources of pronunciation deficiency and suggests ways of improvement to address quality issues. After taking remedy measures, subjective tests on improved system are carried out to validate the exercise. The test results are encouraging.