Signal and feature domain enhancement approaches for robust speech recognition

This paper analyzes the impact of various preprocessing modules to improve the performance of automatic speech recognition system (ASR) in noisy environment. After choosing the state-of-the-art algorithms designed in the signal domain and feature domain, their performances in various noise conditions are thoroughly evaluated. Since the enhancement has been directly made to the features that are actually used for recognition, the feature domain approach is more appropriate than the signal domain approach. Experimental results show that the noise reduction in the feature domain gives the best performance.