语音识别的鲁棒性特征提取方法研究

训练环境和测试环境的不匹配是造成实际情况下语音识别性能下降的主要原因。在深入研究语音识别的噪声环境和Mel域倒谱系数(MFCC)流程的基础上,基于累计分布函数匹配思想,给出了3种通过减小训练环境和测试环境的不匹配度来提高系统在不同环境下适应性的鲁棒性特征提取方法,分析了它们的理论基础、基本算法,并在Aurora2.0数据库上进行了实现,验证了方法的有效性,为实际应用中如何选择语音识别系统提供了参考。

Research on Robust Feature Extracting Methods of Speech Recognition

The mismatch between the training and testing environments is the main reason which reduces the performance of the speech recognition.Based on research on the speech recognition noise environment and the MFCC,this paper presents three robust feature extracting methods,which can improve the system adaptability in different environments by reducing the mismatch of the training and testing environments based on the idea of cumulative distribution function matching,analyzes their theoretical foundations and basic algorithms,implements on Aurora 2.0,and confirms the validity of the methods,it provides reference to the choice of speech recognition system in actual applications.