On the use of DNN Autoencoder for Robust Speaker Recognition

In this paper, we present an analysis of a DNN-based autoencoderfor speech enhancement, dereverberation and denoising.The target application is a robust speaker recognition system.We started with augmenting the Fisher database with artificiallynoised and reverberated data and we trained the autoencoderto map noisy and reverberated speech to its clean version.We use the autoencoder as a preprocessing step for a stateof-the-art text-independent speaker recognition system. Wecompare results achieved with pure autoencoder enhancement,multi-condition PLDA training and their simultaneous use. Wepresent a detailed analysis with various conditions of NIST SRE2010, PRISM and artificially corrupted NIST SRE 2010 telephonecondition. We conclude that the proposed preprocessingsignificantly outperforms the baseline and that this techniquecan be used to build a robust speaker recognition system forreverberated and noisy data.