基于三音子模型连续语音声调识别方法  

Multi-feature Based Tri-phone Model for Tone Recognition of Chinese Continuous Speech

作　　者：魏瑞莹 梁维谦[2] 

机构地区：[1]清华大学微纳电子学系,北京100084 [2]清华大学电子工程系,北京100084

出　　处：《电声技术》2011年第8期34-37,共4页Audio Engineering

摘　　要：作为汉语语音识别的重要组成部分,声调识别具有关键的作用。提出了一种新的基于前后文相关的模型识别方法用以提高汉语连续语音中的识别率。首先介绍用于声调识别的基因轨迹的提取和处理,然后提出6种特征来描述基因轨迹的变化趋势并给出具体的计算公式,利用这些特征并考虑连续语音中前后音节的相关性对基因轨迹造成的变化而建立细分的声调模型,最后基于这种声调模型采用决策树的分类方法进行声调的识别和测试。As an important part of the recognition of Chinese speech, the tone recognition is also a gut issue. In this paper, a new kind of context-dependent tone model is proposed. Firstly, the Fundamental Frequency （FO） extraction is introduced, and then six kinds of features which contain important information about the discrimination between each type of the tone are proposed. Based on these features and the consideration of the syllable correlations in continuous speech, tri-phone tone models are built. Finally, the decision tree is used to recognize the type of the tone.

关 键 词：声调识别 基因轨迹 特征提取 三音子模型 决策树 

分 类 号：TN912[电子电信—通信与信息系统]