A Unified Totally-Data-Driven Prediction of Duration and Pause in TTS

This paper proposes a unified framework for duration and pause prediction in TTS. The framework is based on plain generalized linear model (GLM) under the assumption of Gaussian distribution for duration and logistic GLM under the assumption of Bernoulli distribution for pause. Significant attributes and attribute interactions can be automatically selected to guarantee a good balance between model reliability and goodness of fit. Both the linear attributes and the nonlinear attribute interactions are selected in a totally-data-driven manner rather than intuitively. Speaking rate is introduced as a new attribute, which increases the prediction precision and provides a novel approach to adjust speaking rate when synthesizing as well. GLM is trained by stepwise regression under F-test and Bayes information criterion (BIC). Open test experiments shows that the proposed GLM approach outperforms the decision tree based approach. In addition, the GLM approach is compact and suitable for embedded TTS