F0 analysis and modeling for Cantonese text-to-speech
Eva Yujia Li, Tan Lee, Yao Qian · 2004
This paper presents a study on the control of fundamental frequency (F0) in Cantonese text-to-speech (TTS) systems.The surface F0 contour of an utterance is considered as the combination of tone-related local components and phrase-level long-term variation.A novel method of F0 normalization has been developed to effectively separate them.Statistical analysis is performed for the phrase curves and the tone contours extracted from a large speech corpus, and the results are summarized into regular patterns.These patterns are used as the basic templates in a non-parametric F0 model, from which utterance-level F0 contours can be generated.Perceptual test shows the naturalness of speech naturalness is significantly improved by the new F0 model.The MOS increases by 0.65 over a five-point scale.