利用五度字调模型实现普通话到兰州方言的转换-声学技术.pdfVIP

第 29 卷第 6 期 声 学 技 术 Vol.29, No.6 2010 年 12 月 Technical Acoustics Dec., 2010 利用五度字调模型实现普通话到兰州方言的转换 梁青青，杨鸿武，郭威彤，裴 东，甘振业 (西北师范大学物理与电子工程学院，兰州 730070) 摘要：方言语音的转换对于实现个性化的汉语语音合成系统具有重要的意义。为了实现普通话到兰州方言的转换， 论文利用《方言调查字表》建立了兰州方言的文本语料库，录制了普通话和兰州方言平行的语音语料库。利用五度 字调模型建立了普通话和兰州方言的基频模型，利用统计方法建立了时长转换模型和停顿时长转换模型。在转换时， 通过利用 STRAIGHT 算法修改普通话语音实现普通话到兰州方言的转换。对转换结果的 MOS 评测表明，转换后的 单字平均 MOS 得分 4.17 ，双字平均MOS 得分为 4.22 ，语句的平均MOS 得分为 3.55 。 关键词：语料库；基频特征；时长特征；韵律模型；方言转换 中图法分类号：TP391 文献标识码：A 文章编号：1000-3630(2010)-06-0620-06 DOI 编码：10.3969/j.issn1000-3630.2010.06.014 Mandarin to Lanzhou dialect conversion with five degree tone model LIANG Qing-qing, YANG Hong-wu, GUO Wei-tong, PEI Dong, GAN Zhen-ye (College of Physics and Electronic Engineering, Northwest Normal University, Lanzhou 730070, China) Abstract: Dialect conversion is the one most important aspects of Chinese speech synthesis. A Lanzhou dialect corpus has been built based on word-list in dialectal survey for the conversion of Lanzhou dialect from Mandarin. Speech corpus is recorded with contrastive (Lanzhou dialect vs. Mandarin) recordings. Five degree tone model based prosodic models are built for Lanzhou Dialect by analyzing the differences of pitch, duration and pause duration between Lanzhou dialect and Mandarin. Lanzhou dialect is converted from Mandarin by STRAIGHT algorithm. Subjective experiments demonstrate that the converted monosyllable, disyllable and sentence achieve 4.17, 4.22 and 3.55 of the average mean opinion score (MOS). Key words: corpus; fundamental frequency; duration; prosody model; dialect conversion