基于AdaBoost的汉语方言辨识  

AdaBoost based Chinese dialect identification

作　　者：顾明亮[1] 夏玉果[2] 张长水[3] 杨亦鸣[2] 

机构地区：[1]徐州师范大学物理与电子工程学院,徐州221116 [2]江苏省语言科学与神经认知工程重点实验室,徐州221116 [3]清华大学自动化系,北京100086

出　　处：《东南大学学报（自然科学版）》2008年第4期585-588,共4页Journal of Southeast University：Natural Science Edition

基　　金：国家社会科学基金重点资助项目(01AYY004);江苏省"十五"社科基金资助项目(K3-013);江苏省高校自然科学基金资助项目(99KJB510002);徐州师范大学重大培育资助项目

摘　　要：为了在训练样本受限的情况下,提高汉语方言辨识的效果,提出了一种基于AdaBoost的汉语方言辨识新方法.该方法将GMM与语言模型组成的辨识系统看成一组弱分类器,然后对这组弱分类器所得的分类结果进行加权投票,最终决定汉语方言测试语音的所属类别.实验结果表明:增加GMM或弱分类器的个数,可以有效提高系统的辨识效果;测试语音越长,系统辨识效果越好;当训练样本有限的情况下,采用AdBoost方法比采用ANN方法具有更高的辨识率.In order to improve the performance of Chinese dialect identification under the confined training data, a novel dialect identification method using AdaBoost algorithm is presented. The new method uses the results of a set of ＂poor＂ classifiers, which consist of Gaussian mixture model （GMM） and language models, to vote and produce the final decision. According to experimental results, the following conclusions are obtained：The performance of the system can be improved effectively by increasing the number of GMM and the ＂poor＂ classifiers. The longer the length of test speech is, the higher the identification accuracy of the system is. Using the AdaBoost method can get higher recognition rate than using artifical neural network （ANN） approach under the restricted training data.

关 键 词：ADABOOST算法 高斯混合模型 方言辨识 

分 类 号：TP391.42[自动化与计算机技术—计算机应用技术]