基于子带GMM-UBM的广播语音多语种识别  

Broadcast Speech Language Recognition Based on Sub-Band GMM-UBM

作　　者：李思一[1] 戴蓓蒨[1] 王海祥[1] 

机构地区：[1]中国科学技术大学电子科学与技术系,合肥230026

出　　处：《数据采集与处理》2007年第1期14-18,共5页Journal of Data Acquisition and Processing

摘　　要：提出了一种基于概率统计模型的与语言内容无关的语种识别方法,它不需要掌握各语种的专业语言学知识就可以实现几十种语言的语种识别;并针对广播语音噪声干扰大的特点,采用GMM-UBM模型作为语种模型,提高了系统的噪声鲁棒性;由于广播语音的背景噪声不是简单的全频带加性白噪声,因此本文构建了一种基于子带GMM-UBM模型的多子系统结构的语种识别系统,后端采用神经网络进行系统级融合。本文通过对37种语言及方言的识别实验,证明了子带GMM-UBM方法的有效性。A language recognition method is proposed based on probability-statistical model. It can recognize several decade kinds of languages without professional linguistic knowledge. Aimed at the high noise of the broadcast speech, GMM-UBM is used as the language model to improve the system noise-robustness. And because the background noise of the broadcast speech is not the simply full-band Gaussian white noise, a language recognition system is built based on sub-band GMM-UBM model and subsystems structure by using neural network to fuse different subsystems. Experimental results for recognizing 37 languages and dialects verify the validity of the sub-band GMM-UBM method.

关 键 词：语种识别 语言内容无关 广播语音 子带GMM—UBM 

分 类 号：TN912[电子电信—通信与信息系统]