IMPROVEMENT OF THE SYNTHESIZED SPEECH INTONATION WITH STYLISATION ANDNEURAL-NETWORK LEARNING

A new F0 contour generation model with stylisation and neural network learning is proposed. Our F0 contour consists of (i) global tune, (ii) word pitch bias, (iii) lexical tone, and (iv) syllabic pitch pattern. A preconstructed melodic table takes care of (i), (ii) and (iv) while a neural network deals with (iii). A 5-level subject test shows 0.5 p oint intonation quality improvement compared to our previous model.