Simultaneous Acoustic, Prosodic, and Phrasing Model Training for TTs Conversion Systems

A new integrated model for simultaneous modeling of linguistic and acoustic models, and a training algorithm is proposed. Usually, text-to-speech (TTS) systems based on the hidden Markov model (HMM) consist of text analysis and speech synthesis modules. Linguistic and acoustic model training are performed independently using different training data sets. Integrated model parameters were simultaneously optimized by the proposed training algorithm. The derived algorithm optimizes two model parameter sets simultaneously. Therefore, the appropriate model is estimated because we can directly-formulate the TTS problem in which the speech waveform is generated from a word sequence. We conducted objective evaluation experiments using phrasing and prosodic models to evaluate the effectiveness of the proposed technique.