基于正弦谐波模型的语音变换算法及实现  

Algorithm and implementation of voice transformation technology based on harmonic sinusoidal model

作　　者：王浩[1] 苏巨诗 许胜华[3] 岳振军[3] 

机构地区：[1]解放军理工大学通信工程学院,江苏南京210007 [2]南京军区空军堪察设计院,江苏南京210002 [3]解放军理工大学理学院,江苏南京211101

出　　处：《解放军理工大学学报（自然科学版）》2005年第6期525-530,共6页Journal of PLA University of Science and Technology(Natural Science Edition)

摘　　要：介绍了语音变换的相关技术,分析了利用正弦谐波模型实现语音变换的算法及流程。利用正弦谐波模型对语音进行建模和分解,提取语音的基音频率,利用高斯建模和变换实现语音韵律特征的变换;提取出正弦谐波幅度的后10阶系数,作为语音的频谱特征参数,利用矢量量化和码书映射的方法实现语音频谱特征的变换。提出了一种逐词对应的训练参数对齐方法,给出了具体实现的算法流程。对录制的2段语音利用该算法进行了仿真实验,利用ABX测试对实验结果进行了评估。测试结果显示,该算法得到的变换语音在听觉上有89.3%的概率更接近目标说话人语音。The technology of voice transformation was introduced. The algorithm of voice transformation based on harmonic sinusoidal modal was analyzed. The speech was modeled and decomposed by use of the harmonic sinusoidal modal; the pitch was extracted and the rhythm character of speech was transformed based on gauss model, the last 10 coefficients of harmonic sinusoidal model amplitude were extracted and the spectral feature was transformed based on VQ and codehook mapping method. A technology characteristic of face to face according to word to word was put forward and the algorithm was given. The experiment based on proposed algorithm was applied to two segments speech. The ABX test results indicate that the converted speech is 89.3% similar to that of the target speaker.

关 键 词：语音变换 正弦谐波模型 基音频率 频谱特征 矢量量化 码书映射 

分 类 号：TN912.3[电子电信—通信与信息系统]