Channel Compensation for Speaker Recognition using MAP Adapted PLDA and Denoising DNNs.

: AbstractOver several decades, speaker recognition performance hassteadily improved for applications using telephone speech. Abig part of this improvement has been the availability of largequantities of speaker-labeled data from telephone recordings.For new data applications, such as audio from room microphones,we would like to effectively use existing telephone datato build systems with high accuracy while maintaining goodperformance on existing telephone tasks. In this paper we compareand combine approaches to compensate models parametersand features for this purpose. For model adaptation weexplore MAP adaptation of hyper-parameters and for featurecompensation we examine the use of denoising DNNs. On amulti-room, multi-microphone speaker recognition experimentwe show a reduction of 61% in EER with a combination of theseapproaches while slightly improving performance on telephonedata.