Hybrid Source Modeling Method Utilizing Optimal Residual Frames for HMM-based Speech Synthesis

This paper proposes a new hybrid source modeling method for improving the quality of HMM-based speech synthesis. The proposed method is an extension of recently proposed source model based on optimal residual frame [1]. The source or excitation signal is first decomposed into a number of pitch-synchronous residual frames. Unique variations are observed in the pitch-synchronous residual frames present at the beginning, middle and end regions of excitation signal of a phone. Based on the observation, one optimal residual frame is extracted from each of the beginning, middle and end regions of excitation signal of a phone. The optimal residual frames extracted from every region of excitation signal are separately grouped in the form of decision tree. During synthesis, for every phone, three optimal residual frames are selected from three decision trees based on target and concatenation costs. Using three optimal residual frames, the excitation signal of a phone is constructed. The proposed hybrid source model is used for synthesizing speech under HTS framework. Subjective evaluation results indicate that the proposed source model is better the two existing source modeling methods.