Tree-based modelling for speech synthesis