应用说话人聚类技术改善语言辨识系统识别率  

Improving Automatic Language Identification by Speaker Clustering Methods

作　　者：张强[1] 屈丹[1] 侯风雷[1] 王炳锡[1] 

机构地区：[1]中国人民解放军信息工程大学,河南郑州450002

出　　处：《电声技术》2007年第3期44-48,共5页Audio Engineering

摘　　要：说话人差异是影响语言辨识系统性能的一个重要方面。采用说话人聚类技术对训练数据进行了预分类,以此为基础训练各种语言模型,得到聚类后的语言辨识系统。高斯混合模型、高斯混合模型-全局背景模型和遍历隐马尔可夫模型3种语言模型的实验证明,这种方法是有效的,它普遍地减小了说话人差异对语言辨识的影响,提高了语言辨识率。The speaker difference is an important factor which affects the performance of the automatic language identification system, The speaker clustering method is applied to classify training-data in advance, By training the language model by these classified data, a new LID system is got. Experiments based on the reference system including GMM（Gaussian Mixed Model）, GMM-UBM（Gaussian Mixture Model-Universal and Background Model） and HMM （Hidden Markov Model） show this method is effective, that the effect of speaker difference on LID performance is decreased, and that the language identification rate is improved in general.

关 键 词：说话人聚类 高斯混合模型-全局背景模型 遍历隐马尔可夫模型 

分 类 号：TN912[电子电信—通信与信息系统]