Speaker Verification with Application-Aware Beamforming

Multichannel speech processing applications usually employbeamformers as means of speech enhancement through spatialfiltering. Beamformers with learnable parameters requiretraining to minimize a loss function that is not necessarilycorrelated with the final objective. In this paper, we presenta framework employing recent neural network based generalizedeigenvalue beamformer and application-specific modelthat allows for optimization of beamformer w.r.t. target application.In our case, the application is speaker verificationwhich utilizes a speaker embedding (x-vector) extractorthat conveniently comes with desired loss. We show thatapplication-specific training of the beamformer brings performanceimprovements over a system trained in the standardway. We perform our analysis on the recently introducedVOiCES corpus which contains multichannel data and allowsus to modify the evaluation trials such that enrollment recordingsremain single-channel and test utterances are multichannel.