Real-Time Reverb Reduction for Improved Automatic Speech Recognition in Far-Field

In the paper, methods of real-time reverb reduction based on Generalized Weighted Prediction Error (GWPE) were presented. It was shown that usage of the proposed audio processing routines highly improve the accuracy of Automatic Speech Recognition (ASR) system namely word error rates (WERs) are reduced 11.36% when the user stands 5 meters from the microphone array. The obtained results are close to the ones that are achieved by the offline GWPE implementation (12.06%). Thanks to optimizations and parameters tuning, computational complexity of the proposed realization of GWPE was highly reduced and it achieves RTFs lower than 1.0 (computation time is shorter than signal duration) when using one core of CPU.