摘要
语音识别系统在实用环境中的鲁棒性是语音识别技术实用化的关键问题。鲁棒性研究的核心问题是如何解决实用环境语音特征和模型与干净环境语音识别系统的失配问题,这涉及到噪声补偿、信道适应、说话人自适应等关键技术。文章综述了语音识别鲁棒性技术研究的主要方法、原理及研究现状,分析了实用环境语音识别中声学模型和语言模型的适应技术,并展望了近期语音识别实用化技术发展的研究方向。
The robustness of speech recognition system in real environment is the key problem in the application of speech recognition technique.The kernel problem is how to overcome the mismatching of acoustic feature and model between real and clean environments,which involves the important techniques on noise environment compensation, channel adaptation and speaker adaptation etc.This paper reviews the main methods,principles and ad-hoc research techniques of robust researches,analyzes the adaptation techniques of acoustic model and language model in real environment,and proposes the recent prospect of research interests in real-world environment speech recognition.
出处
《计算机工程与应用》
CSCD
北大核心
2006年第24期7-12,共6页
Computer Engineering and Applications
基金
国家863高技术研究发展计划资助项目(编号:2001AA114071)
中国博士后科学基金
清华大学博士后基金
北京航空航天大学"985二期""复杂系统与空间物质结构科技创新平台"项目资助
关键词
语音识别
实用环境
稳健性
信道适应
说话人自适应
speech recognition,real-world environment,robustness,channel adaptation,speaker adaptation
