| Summary: | 碩士 === 國立成功大學 === 電機工程學系 === 102 === In this thesis, a vivid storytelling system is proposed for making audience immerse deeply into the story and also enhancing the interest of storytelling. For conveying emotion more vividly, the nonverbal paralinguistic cues like affective prosody and nonverbal vocalizations (laughter, shout, weeping and etc.) play an important roles in storytelling. In addition, role-playing element such like speaker style applied into storytelling can improve the dramatic impression of story. For this purpose, this storytelling system is with two artifices: 1) convey emotions by nonverbal emotional cues 2) convert the voice of storyteller to the specific story role.
In order to generate the nonverbal emotional cues, we synthesis expressive speech by HMM-based TTS and insert nonverbal emotional vocalization (NEV) by a generation module consisted of emotion classification based on SVM and a template-based NEV prediction model. In addition, for generating different story role’s voice, the speaker conversion based on formant warping is used to conquer the problem of recording a heavy amount of multiple speaker’s expressive style speech corpora.
The experimental result shows that our storytelling system can achieve a MOS of 4.43, it indicates that the proposed storytelling is more near to real storyteller except a slight unnaturalness. Furthermore, the storytelling with emotional expression and role-playing is more natural for audience, and captures more audience’s attention.
|
|---|
