基于音素分类的汉语语声转换算法.PDFVIP

31 1 ( ) Vo.l 31 No. 1 2011 2 Journal of Nanjing University of Posts andTelecomm un ications( Natural cience) Feb. 2011 1 1 2 李燕萍 , 张玲华 , 丁 辉 1. , 210003 2. , 314001 : 提出了一种基于音素分类的语声转换算法根据汉语语音学理 , 分析汉语语言结构, 通过帧特征分析, 构建了汉语元音映射表,指导音素的分类, 然后对每类音素进行独立的特征参数转换, 解决了传统 GMM 转换算法 的过平滑问题, 避免了训练与测试阶段的语料音素之间的不均衡, 从而提高语音的可懂度与自然度客观评测结 果表明, 基于音素分类的语声转换算法与 GMM 算法相比, 谱距离比值平均降低了 9. 31% ; 主观听力测试表明, ABX 测试结果提高了 10. 36% , 同时MO 得分平均提高了 29. 33% : 语声转换; 音素分类; 汉语元音映射; 高斯混合模型 : TN912. 3 : A : 2011) A NovelAlgorithm forChineseVoiceConversion Based on PhonemeClassification 1 1 2 LIYanping, ZHANG Linghua, ING Hui 1. College ofTelecommunications Inform ation Engineering, Nanjing University of Posts andT elecomm unications, Nanjing 210003, China 2. Co llege ofM athem atics and Inform ation Engineering, Jiaxing University, Jiaxing 314001, China Abstract: A novel algorithm for Chinese voice conversion based on phoneme classification is proposed in th is paper. The analysis of Ch inese linguistic structure provides the construction of vow elm apping theory w hich guides the classification of phonemes. The proposed method mi plements feature parameters conver sion for each category phonem e independently. It deals w ith the spectral smooth ing of trad itional GMM ( Gaussian M ixtureModel) and prevents phonem e mi balance betw een training and testing materials and then mi proves voice intelligib ility and naturalness. mi ilarity to the target voice spectral is evaluated in an objective test and