Prosody Modelling for TTS Systems Using Statistical Methods

The main drawback of older methods of prosody modelling is the monotony of the output, which is perceived as uncomfortable by the users, especially when listening to longer passages. The present paper proposes a prosodic generator designed to increase the variability of synthesized speech in reading devices for the blind. The method used is based on text segmentation into several prosodic patterns by means of vector quantisation and the subsequent training of corresponding HMMs (Hidden Markov Models) on F0 parameters. The path through the model's states is then used to generate sentence prosody. We also tried to utilize morphological information in order to increase prosody naturalness. The evaluation of the quality of the proposed prosodic generators was carried out by means of listening tests.