Multi-level prosody and spectrum conversion for emotional speech synthesis

Emotional speech can be synthesized by converting prosodic and spectrum features in neutral speech. This paper propose a multi-level prosody conversion method, it converts three prosodic features of F0, short-time energy and speaking rate in syllable, prosodic word and sentence level sequentially. The F0 and speaking rate is modeled by Gaussians, and energy is modeled by Gamma distribution respectively. The experiments both of objective and subjective evaluation test show that proposed method is effective for emotion conversion of speech.