基于自回归模型的加性噪声环境稳健语音识别  

Autoregressive model-based robust speech recognition in additive noise environment

作　　者：刘敬伟[1] 王作英[1] 肖熙[1] 

机构地区：[1]清华大学电子工程系,北京100084

出　　处：《清华大学学报（自然科学版）》2006年第1期50-53,共4页Journal of Tsinghua University(Science and Technology)

基　　金：国家"八六三"高技术项目(2001AA114071);中国博士后科学基金项目;清华大学博士后基金项目

摘　　要：为提高噪声不平稳或不可估的情况下语音识别的稳健性,提出了利用自回归模型和短时平稳性假设,估计干净与噪声环境的语音数据,建立相应的语音识别模型,以达到抗噪效果的稳健语音信号处理方法。在N o iseX-92的4种噪声环境(w h ite,babb le,vo lvo,destroyer eng ine)从0到20 dB的不同信噪比下的“863”大词汇连续语音标准数据库的平均识别结果表明,该方法能够使得基于段长分布的隐M arkov模型的语音识别系统在25候选时声学层的音节相对错误率下降达到10.85%以下,同时相对正确识别率上升12.13%。The robustness of speech recognition in a non-stationary or unknown additive noise environment is intprovecl by a speech signal processing method based on an autoregressive （AR） model using the short time stationary assumption. The method caplores the model-based speech information in both clean and noisy speech signals using a hidden Markov model （HMM） With the AR-estimated speech signal. Experimental results using the standard ＂863＂ large vocabulary continuous speech database with four noise （white, babbitt, volvo, destroyer engine） from NoiseX-92 at signal-to-noise ratios （SNR） from 0 to 20 dB show that the AR model processing reduces the acoustic level relative syllable error rate by 10, 9% and improves the relative syllable recognition rate by 12.1% when using 25-candidate selection in the duration distribution-based hidden Markov model speech recognition system.

关 键 词：语音识别 稳健性 自回归模型 段长分布 隐含Markov模型 

分 类 号：TN912.34[电子电信—通信与信息系统]