PROBABILISTIC ACOUSTIC MODELLING FOR PARAMETRIC SPEECH SYNTHESIS

This thesis proposes a new probabilistic model and speech parameter generation method for statistical parametric text-to-speech synthesis. These are designed to ad-dress drawbacks with the conventional approach to statistical parametric speech syn-thesis, in which a form of hidden Markov model (HMM) is used for modelling and one of two methods, known as standard speech parameter generation and parameter generation considering global variance, is used for speech parameter generation. The new probabilistic model is a form of autoregressive HMM, and addresses an inconsistency in the treatment of the dynamics of speech parameter sequences present in the conventional approach. Compared to previous attempts to address this inconsist-ency such as the trajectory HMM, our model has the advantage of supporting efficient parameter estimation using expectation maximization and decision tree clustering. The new speech parameter generation method is informed by a mathematical ana-lysis of parameter generation considering global variance and an investigation into why it sometimes introduces artifacts into the synthesized speech. The proposed method is as fast as standard speech parameter generation but improves naturalness as much a