基于多特征组合的普通话塞音识别

Mandarin plosive recognition based on multi-feature combinations

针对汉语塞音发音易混淆、变化速率快等不易识别的问题,提出在语音和声学特征基础上,加入其他特征参数来提高汉语塞音的识别性能。提取的参数包括嗓音起始时间(VOT)、音轨方程、发音器官运动轨迹位移、速度和加速度的运动学特征,并将提取的声学和运动学特征进行融合,形成不同的特征组合;再分别对特征组合进行主成分分析(PCA)和信息熵计算;最后通过SVM识别网络,测试特征组合的识别性能。测试结果显示,通过PCA后特征组合识别率排名Top-10的组合与熵计算后的特征组合排名一致,表明特征组合识别塞音的稳定性;且与单组特征相比,Top-10特征组合识别率都有提高,PCA后其识别率最高达到97.45%。

In allusion to the problems of easy confusion,fast change rate,and other recognition difficulties during the Chinese plosive pronunciation,a method of adding other feature parameters on the basis of the phonetic and acoustic features is proposed to improve the recognition performances of Chinese plosives. The extracted parameters include voice onset time(VOT),audio track equation,motion track displacement of speech organs,and kinematical features of velocity and acceleration. The extracted acoustic and kinematic features are merged to form different feature combinations. The principal component analysis(PCA)and information entropy calculation are conducted respectively for the feature combinations. The recognition performances of the feature combinations are tested by means of the SVM recognition network. The test results show that the combinations whose recognition rates rank in TOP-10 in the feature combinations after PCA are consistent with those of the ranked feature combinations after entropy calculation,which demonstrates the stability of plosive recognition got by the feature combinations,and in comparison with the single feature,the recognition rates of TOP-10 feature combinations are all improved,and the recognition rate after PCA can reach as high as 97.45%.