Linguistic Processor Training on Speaker Data for Unit Selection Text-to-Speech

This paper describes an approach to synthesizing personalized speech while maintaining not only speaker voice but also speaker pronunciation peculiarities. Personalization is realized by means of pronunciation models trained on speaker data contained in his/her speech database. Untrained models allow to synthesize speech in neutral normative style. On the segmental level, the transcription model is used. On the prosodic level, models for phrasing, intonation, pause and phoneme duration are used. These prosodic models are derived from comparative acoustic-phonetic study of different speakers’ data contained in several speech corpora and databases. Personalizing of pronunciation models is carried out during the off-line training of linguistic processor using a speech database annotation. During the on-line speech synthesis mode, personalized pronunciation models are used by the linguistic processor to generate speaker specific target specification of input text.