engleski

Prosodic Modelling for Croatian Speech Synthesis

In order to include prosody into the text to speech (TTS)systems prosody knowledge needs to be acquired, represented and incorporated. Two main features of prosody important for modelling prosody for TTS systems are duration and F0 contour. There are various approaches to modelling those features and they can be categorized into three main groups: rule based, statistical and minimalistic. Some of the best known approaches to duration acquiring are Klatt’ s model, classification and regression trees (CARTS) and neural networks and to F0 modelling TOBI, Fujisaki and Tilt. A procedure for automatic intonation event detection on Croatian texts based on the Tilt model was evaluated in terms of Root Mean Square Error values for generated F0 contours.

prosody modelling, speech synthesis, TTS, duration models, F0 contour models, prosodic characteristics of Croatian

nije evidentirano