Model Compensation Using Robust Features for Robust Speech Recognition

Gang Wang
2003-01-01
Abstract: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.
What problem does this paper attempt to address?