Dereverberation and Beamforming in Robust Far-Field Speaker Recognition

This paper deals with robust speaker verification (SV) in farfieldsensing. The robustness is verified on a subset of NISTSRE 2010 corpus retransmitted in multiple real rooms of differentacoustics and captured with multiple microphones. Weexperimented with various data preprocessing steps includingdifferent approaches to dereverberation and beamforming appliedto ad-hoc microphone arrays. We found that significantimprovements in accuracy can be achieved with neural networkbased generalized eigenvalue beamformer preceded byweighted prediction error dereverberation. We also exploredthe effect of data augmentation by adding various real or simulatedroom acoustic properties to the Probabilistic Linear DiscriminantAnalysis (PLDA) training dataset. As a result, wedeveloped a speaker recognition system whose performanceis stable across different room acoustic conditions. It yields41.4% relative improvement in performance over the systemwithout multi-channel processing tested on the cleanest microphonedata. With the best combination of data preprocessingand augmentation, we obtained a performance close to the onewe achieved with the original clean test data.