DNN FEATURE COMPENSATION FOR NOISE ROBUST SPEAKER VERIFICATION

The speaker verification (SV) task has been an active area of research in the last thirty years. One of the recent research topics is on improving the robustness of SV system in challenging environments. This paper examines the robustness of current state of the art SV system against background noise corruptions. Specifically, we consider the scenario where the SV system is trained from noise free speech and tested on background noise corrupted speech. To improve robustness of the system, a deep neural networks (DNN) based feature compensation is proposed to enhance the cepstral features before the evaluation. The DNN is trained from parallel data of clean and noise corrupted speech which are aligned in the frame level. The training is achieved by minimizing the mean square error (MSE) between the DNN's prediction and the target clean features. The trained network could predict the underlying clean features when given noisy features. Results on the benchmarking SRE 2010 female core task show that by using DNN based feature compensation, the equal error rate (EER) can be reduced in most of the times even when the test noise is unseen during DNN training. The relative EER reduction usually is in the range of 3% to 26%.