Dereverberation and Beamforming in Robust Far-Field Speaker Recognition

speaker verification, beamforming, dereverberation, autoencoder

This paper deals with robust speaker verification (SV) in farfield sensing. The robustness is verified on a subset of NIST SRE 2010 corpus retransmitted in multiple real rooms of different acoustics and captured with multiple microphones. We experimented with various data preprocessing steps including different approaches to dereverberation and beamforming applied to ad-hoc microphone arrays. We found that significant improvements in accuracy can be achieved with neural network based generalized eigenvalue beamformer preceded by weighted prediction error dereverberation. We also explored the effect of data augmentation by adding various real or simulated room acoustic properties to the Probabilistic Linear Discriminant Analysis (PLDA) training dataset. As a result, we developed a speaker recognition system whose performance is stable across different room acoustic conditions. It yields 41.4% relative improvement in performance over the system without multi-channel processing tested on the cleanest microphone data. With the best combination of data preprocessing and augmentation, we obtained a performance close to the one we achieved with the original clean test data.