Intonation processing for TTS using stylization and neural network learning method

We propose a new model for synthesizing fundamental frequency (F0) contours using a stylisation and a neural network learning method. The F0 contour is described as the superposition of 4 layered features; global tune, word pitch bias, lexical tone, and the syllabic pitch pattern. We firstly stylize the F0 contour of speech material, and analyse stylised data by a statistical approach according to grammatical attributes. We then construct a melodic table, and train lexical tone with a neural network. Finally we develop the intonation generation rules for TTS conversion. This model produces a good neutral declarative intonation, and there is little difference between synthesized speech with the original F0 contour and that with the rule generated content when tested with our TD-PSOLA synthesizer