基于语音结构化模型的数字语音识别  

Digital recognition based on acoustic universal structure

作　　者：姜莹[1] 俞一彪[1] 

机构地区：[1]苏州大学电子信息学院,江苏苏州215006

出　　处：《计算机工程与设计》2012年第4期1482-1485,1490,共5页Computer Engineering and Design

基　　金：北京市"现代信息科学与网络技术"重点实验室暨铁道部"铁道信息科学与工程"开放实验室开放基金项目(XDXX1006)

摘　　要：提出一种新的基于语音结构化模型的语音识别方法,并应用于非特定人数字语音识别。每一个数字语音计算倒谱特征之后提取语音中存在的对说话人差异具有不变性的结构化特征——全局声学结构(acoustical universal structure,AUS),并建立结构化模型,识别时提取测试语音的全局声学结构,然后与各数字语音的结构化模型进行匹配。测试了少量语料训练下的识别性能并与传统HMM(hidden Markov model)方法进行比较,结果表明该方法可以取得优于HMM的性能,语音结构化模型可以有效消除说话人之间的差异。Recently, a novel acoustic representation of speech is proposed, called the acoustic universal structure, and is applied to a speaker-independent digital recognition system. After each digit utterance is converted into a sequence of distributions, the acoustic universal structure which can well remove the speaker variations is extracted to establish the structural model. In recog nition stage, the acoustic universal structure is extracted from the test utterance, and then match with each digit＇s structural model. The candidate digit show the maximum likelihood score is the result of recognition. The experiments are conducted under the condition of training data only contains a single digit speech, and the performance of this method is compared with the tradi- tional HMM. The results show the proposed method has better performance than HMM, and can discard different speaker varia bility effectively.

关 键 词：语音结构化模型 数字识别 隐马尔可夫模型 说话人差异 巴氏距离 

分 类 号：TN912.34[电子电信—通信与信息系统]