Towards Robust Distant-Talking Automatic Speech Recognition in Reverberant Environments

In distant-talking scenarios, automatic speech recognition (ASR) is hampered by background noise, competing speakers and room reverberation. Unlike background noise and competing speakers, reverberation cannot be captured by an additive or multiplicative term in the feature domain because reverberation has a dispersive effect on the speech feature sequences. Therefore, traditional acoustic modeling techniques and conventional methods to increase robustness to additive distortions provide only limited performance in reverberant environments.

Based on a thorough analysis of the effect of room reverberation on speech feature sequences, this contribution gives a concise overview of the state of the art in reverberant speech recognition. The methods for achieving robustness are classified into three groups: Signal dereverberation and beamforming as preprocessing, robust feature extraction, and adjustment of the acoustic models to reverberation. Finally, a novel concept called reverberation modeling for speech recognition, which combines advantages of all three classes, is described.