多层次韵律与短时谱同步变换的情感语音合成.pdf

多层次韵律和短时谱同步变换的情感语音合成中文摘要 I 多层次韵律和短时谱同步变换的情感语音合成中文摘要在日常生活中,声音包含了表示文本内容的语义信息,而且也会传递一些情感信息。对于同一句话,如果说话人说话方式不同,听者所获得的信息也会不同。语音的情感转换,就是在语义相同的情况下,实现声音在不同情感间的转换。因此,情感转换是具有表现力的语音合成的重要研究方向。为了能够合成出高质量的情感语音,本文使用了一种多层次韵律和短时谱同步变换的情感合成方法。通过多层次的方法对高兴、生气、悲伤和中性这4种情感语音建立相应的韵律模型。在此基础上,训练得到中性语音与情感语音之间的映射关系,完成韵律转换。然后,再结合短时谱的转换,运用合成工具(STRAIGHT)最后合成有明显情感倾向的情感语音。对转换语音做ABX和MOS测评,结果表明多层次的方法明显改善了情感转换效果。同时,对于合成的情感语音进行谱失真检测,检测结果表明,相对于只对音节进行转换的方法,本文对于高兴、愤怒和悲伤的转换结果分别提高了2%、4%和6%。关键词:多层次韵律;短时谱转换;高斯混合模型;情感语音合成作者:王泽勋指导老师:俞一彪 Abstract Multi-level Prosody and Short-term Spectrum Transform for Emotional Speech Synthesis II Multi-level Prosody and Short-term Spectrum Transform for Emotional Speech Synthesis Abstract In daily life, the speech not only contains the meaning of the text content, but also delivers some emotional information. For the same sentences, if the stylethat speaker expresses is different, the information listener get will also be different. Emotional speech conversion, that is, under the condition of the same text, realizing voice conversion between different , the research of emotion transformation plays a significant role in expressivespeech synthesis. In order to synthesizehigh quality of emotional speech, this paper usea multi-level emotional prosodyand short-time spectrum transform for emotional speech synthesis method. By using the method ofmulti-level, we build a corresponding prosodic model for happiness, anger, sadness and neutral speech. Based on it, we realize the prosody conversion after training the mapping relationship between neutral and emotional speech, and then complete the end, combined with the short-term spectrum transformation, weuse STRAIGHT to synthesize obvious emotional speech. In this paper, thesubjective evaluation method of MOSand ABX is used to test the converted emotional speech andtheresults