基于指数加权的基音平滑算法  

Method of pitch smoothing based on exponential weighting

作　　者：屈小刚[1] 蒋保臣[1] 

机构地区：[1]山东大学威海分校信息工程学院,山东威海264209

出　　处：《计算机工程与设计》2006年第17期3265-3266,3308,共3页Computer Engineering and Design

摘　　要：提出了一种用于语音合成的语音片断基音平滑技术。在基于波形拼接的语音合成中,一般使用TD-PSOLA算法进行基频和时长的修改,但是用传统的TD-PSOLA算法进行的基频修改是针对片断整体而言,所以仍然不能很好的解决语音合成中的拼接单元之间的基频不连续问题,特别是在片断接合处。由于基元片断提取自不同语境的语料,合成语音听起来明显感觉到音高的不自然。对传统的TD-PSOLA算法进行了改进,以基音周期为间隔对语音片断信号进行分帧,通过指数加权相应帧的方法来进行平滑处理,经听音测试,较好的解决了拼接片断间的不连续现象。A method of pitch smoothing for unit segments is presented. In speech synthsis based on waveform concatenative synthesis, time domain pitch synchronous overlap add （TD-PSOLA） algorithm is often used to modify pitch and duration. However, conventional TD-PSOLA algorighm works on whole segment in pitch smoothing, it is still poor to resolve discontinuity between unit segment, particularly at the joint. Because the unit is selected from different context corpus, so it is felt obviously that discontinuous pitch appears as if synthetic speech comes from different intonation. Smooth pitch is tried to smooth according to context by weighting exponential function on each-frame splitted by pitch mark. The listenning test shows that this method well resolve discontinuity problem at the joint.

关 键 词：基音平滑 指数加权 时域基音同步叠加算法 语音合成 基频曲线 基音标记 

分 类 号：TN912.33[电子电信—通信与信息系统]