Robust speech recognition by selecting mel-filter banks

摘要

Mel-filterbank energies is a key feature that is widely employed in automatic speech recognition (ASR) system. It arises from a sub-band spectrum typically. But when the noise exists in the background, Mel-filterbank energies can not be easy to estimated accurately. In this paper, the fact that the trajectories of not only "traditional" log Mel-filterbank energies, but also its delta parameters can be influenced by noise will be theoretically analyzed. As a result, log Mel-filterbank energies and their delta parameters can not be calculated correctly. In this paper, we propose to remove those severely contaminated Mel-filterbank features and only keep those variations which perform better in the speech remained. We demonstrate the effectiveness of this novel operation through speech recognition experiments conducted on the Aurora-2 database.