【作者】 张树森； 范京； 吴聪；

【Author】 Zhang Shusen,Fan Jing,Wu Cong Dept.of Information & Communication Eng.,Beijing Information Technology Institute

【机构】 北京信息工程学院信息与通信工程系；

【摘要】 汉语语调参数是汉语音节的重要特征,中心削波法是提取语音基音并进一步进行汉语语调识别的有效方法。由于实用中一般人语音的音调特征差别很大,例如不同的语速;男女基音的不同;每个人发音特点的不同;连续语音中前后关联语音的不同等,致使目前各方法对汉语语调的识别率仍然较低。本文在四个方面改进了普通中心削波法。一是采用接续帧定域技术改进了基音变化的连续性,二是对基音的峰的突出特征进行质量评估,并应用于语调的辅助识别,三是采用了新的语调参数模型结构,使训练集包括识别时可能遇到的所有发音样本,可以加强对各种特征变化的自适应能力;四是在语调特征参数上改进了普通中心削波法,使用了L1相关技术。另外,对起始帧采用了综合估计技术,从而明显提高了自适应汉语语调识别的正确率。本文从原理上进行了讨论并用实际的测试结果证实了新方法的有效性。更多还原

【Abstract】 Centric clipping is an effective method to detect speech signals pitch and to recognize Chinese sound tone.In practice,the feature of the sound tone of person has a great difference;recognition correct rate to Chinese sound tone by many methods is still very low.In this paper,a new improvement for centric clipping is presented from four aspects:the first one is improvement of continuity of the speech pitch by limiting change range of successive frame;the second one is evaluation of quality of prominent feature for peak value of speech pitch.The third one is enhancement of self-adaptive ability to all kinds of the characteristic variations by using a new model structure of tone parameters.The last one is improvement of general centric clipping and application of L1 correlative technique on feature parameter of the sound tone.In addition,the authors use integrated evaluation technique with the start frame.This paper discusses the new method and conform the efficiency by practical test results.更多还原