The Uncertainty in Prosody of Natural Speech and Its Application in Speech Synthesis
Min Chu · Zhongwen xinxi xuebao · 2004
This paper explores the uncertainty of prosody in a speech corpus, which contains two read versions of 1000 sentences by a professional voice talent under the same linguistic and affective planning. It is found that corresponding prosodic features in the two versions change in a rather wide range. The scope of local variations can be as large as 45-50% of the overall variation range of a speaker. Based on such observation, this paper proposes a minimum error rate criterion (MERC) to replace the traditional maximum correct rate criterion in prosody generation. Furthermore, this paper proposes an approach to integrate the MERC into the unit selection algorithm. Among all instances of a speech unit, those that have the lowest possibility to result unnatural prosody are picked out first, and then the most suitable path is selected from all prosodic equivalent candidates under the smoothest criterion to assure the smoothest concatenation of all units on this path.