基于隐马尔可夫模型和高斯混合模型结合的声音转换方法  

Voice Conversion with the Combination of HMM and GMM

作　　者：岳振军[1] 邹翔[2] 王浩[1] 

机构地区：[1]解放军理工大学理学院,南京211101 [2]解放军理工大学通信工程学院,南京210007

出　　处：《数据采集与处理》2009年第3期285-289,共5页Journal of Data Acquisition and Processing

摘　　要：针对隐马尔可夫模型较强的语音信号表征能力和高斯混合模型良好的声音转换效果,提出了一种了隐马尔可夫模型和高斯混合模型相结合转换线谱频率的方法,给出了理论推导和算法流程,并利用高斯建模实现了韵律特征的转换。利用所述算法对录制的两段语音进行了仿真实验,转换语音有较好的自然度和清晰度,ABX测试结果显示,文中算法得到的语音在听觉上有90.2%的概率更接近目标说话人语音。According to hidden strong representation capability of Markov model （HMM） speech signal and better conversion effect Gaussian mixture model （GMM） ,an approach for line-spectrum frequency transformation using HMM and GMM is presented, and the theoreti- cal derivation and the flow diagram of the algorithm are offered. Then, Gaussian model is introduced to achieve the prosodic feature transformation. The experiment is applied on two segment speech. The experimental result shows that the converted speech has good naturalness and articulation. The ABX test indicates that the converted speech is 90.2% similar to the that of the target speaker.

关 键 词：声音转换 线谱频率 隐马尔可夫模型 高斯混合模型 主观评价 

分 类 号：TN912.3[电子电信—通信与信息系统]