基于高斯混合模型的非母语说话人口音识别  

Accent Identification of Non-native Speaker Based on Gaussian Mixture Models

作　　者：赵征鹏[1] 杨鉴[1] 

机构地区：[1]云南大学信息与电子科学系,昆明650091

出　　处：《计算机工程》2005年第6期148-150,共3页Computer Engineering

基　　金：国家自然科学基金资助项目(60265001)

摘　　要：针对云南境内白族、纳西族、傈僳族 3 个典型的少数民族及汉族普通话语音,采用了高斯混合模型来训练每个民族的口音模型,并用少量的测试语音来获得较为满意的口音分类识别率,目的在于探索降低非母语口音话者语音识别错误率的有效途径。该文通过实验给出,对云南民族口音汉语普通话口音识别,当混合数为 16,语音特征采用 39 维 MFCC 及其一阶、二阶差分参数时,口音识别正确率可达90.83%。This paper concerns mandarin accents of three typical minorities: Bai, Naxi, Lisu, and Han nationality in Yunnan province. Gaussian mixture models (GMMs) are used to train models for each accent, and some small testing utterances are used to achieve a satisfied accent identification rate. The aim of the experiment is to find out an effective method which can decrease the error rate of non-native speech recognition. After some experiments, this paper shows that for accents identification of the minorities in Yunnan, if the components of GMM is 16, the speech feature is 39-dimension MFCC and its first-order and second-order difference parameters, the correct rate of the accent identification can reach 90.83%.

关 键 词：高斯混合模型(GMM) 口音识别 非母语口音 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]