Statistical parametric speech synthesis using a hidden trajectory model

A novel spectral modeling method for statistical parametric speech synthesis using a hidden trajectory model (HTM) is presented in this paper. An HTM is a structured generative model with a two-stage implementation. First hidden formant trajectories are generated from time-aligned formant target sequences by a bidirectional filter. This target-filtering model could provide a correlation structure across temporal frames and describe the effect of co-articulation on speech signals efficiently. Then the observed cepstral features are constituted by a formant-related component and a residual component. The formant-related component is predicted from hidden formant trajectories using a nonlinear and analytical function, and the prediction residuals are modeled by context-dependent Gaussians. In this paper, we apply HTM-based acoustic modeling to speech synthesis and investigate the effectiveness of this method in improving the naturalness and controllability of synthetic speech. Experimental results show that this proposed method can improve the accuracy of spectral feature prediction and the naturalness of synthetic speech compared with the conventional HMM-based method, especially for the conditions where the amount of training data is limited. Furthermore, this method can achieve effective controllability on vowel quality and formant sharpness of synthetic speech by conveniently manipulating the distribution parameters for the phone-dependent targets of formant frequencies and bandwidths.