An overview of Microsoft’s Whistler text-to-speech system

The data-driven approach can significantly facilitate the process of creating text-to-speech (TTS) systems for a new language, a new voice, or a new style. As such, Whistler TTS engine was designed to benefit from automatically constructed model parameters. Efforts to improve Whistler with the use of additional training data and better learning algorithms that make full use of these data will be reviewed. Training data have been augmented for a number of speakers. To better use these data, the hidden Markov model speech recognition system has been used to segment the training corpora and select more representative acoustic units. The classification and regression tree was used for both grapheme to phoneme conversation and unseen triphone generalization. Speech signal reconstruction was based on the mixed excitation source-filter model that leads to better compression of the acoustic inventory. A number of ways to smooth the spectral parameters were also studied to minimize the concatenation distortion. To improve automatically extracted prosodic templates, the learning process was refined with an analysis-by-synthesis approach. However, the coverage remains a challenge for the data-driven approach to make Whistler produce synthetic speech that resembles the original speaker. This is especially true for the prosody model.