一种基于鲁棒特征的模型补偿噪声语音识别方法

针对抗噪声语音特征技术和基于MFCC特征的模型补偿技术在低信噪比时识别率不高的缺点，将抗噪声语音特征和模型补偿结合起来，提出了一种基于单边自相关序列(One—sided autocorrelation，OSA)MFCC特征的模型补偿噪声语音识别方法，以提高语音识别系统在低信噪比时的性能。对0～9十个英文数字和NOISEX92中的白噪声、F16噪声和FACTORY噪声的识别实验结果表明．本文提出的识别方法可以有效地提高OSA—MFCC识别器在噪声环境中的识别率，并且在低信噪比时其性能明显优于经过相同补偿处理的MFCC识别器。

Model Compensation Using Robust Features for Robust Speech Recognition

Robust speech features and model compensation techniques using MFCCs usually fail to achieve high recognition rates when SNR levels are low. To improve the performance of speech recognizer in low SNR conditions, two techniques are combined together and a new model compensation recognition scheme is proposed using the one-side autocorrelation MFCCs. The recognition scheme is evaluated by a 0~9 isolated words, with white noise, F16 noise and FACTORY noise of NOISEX92 adding to the clean speeches at different SNRs. Experimental results show that the recognition scheme can efficiently improve the performance of the OSA-MFCC recognizer in noisy environments and outperform the MFCC recognizer with the same compensation technique in low SNR levels.