Optimal residual frame based source modeling for HMM-based speech synthesis

This paper proposes a method for modeling the excitation signal to improve the quality of HMM-based speech synthesis system (HTS). Single optimal residual frame which closely relates to all frames of phone is chosen to represent the entire residual signal of the phone. Optimal residual frames of all phones present in the speech corpus are efficiently grouped based on positional and contextual features to form a decision tree of clusters. During synthesis, suitable residual frames are picked up from the leaf of decision tree based on certain selection criteria. From each optimal residual frame, the entire residual signal corresponding to the phone is generated based on pitch and intensity contours obtained from HMMs. Subjective evaluation results show that the proposed excitation model can significantly improve the quality of HTS compared to traditional pulse excitation.