Investigation of DNN based Feature Enhancement Jointly Trained with X-Vectors for Noise-Robust Speaker Verification

In this paper, we investigate the deep neural network (DNN) based feature enhancement as the denoising frontend of the x-vector speaker verification framework in noisy environments. Firstly, the feature enhancement DNN (FE-DNN) learns the mapping function from the noisy to the clean corpora on the frame-level acoustic feature domain, and then the x-vector network (XvectorNet) is trained on top of the enhanced features. Finally, the separately trained FE-DNN and the XvectorNet are serially concatenated and jointly trained under the supervision of cross-entropy loss. In addition, we adopt the logistic margin softmax layer for training the XvectorNet in order to obtain more discriminative speaker embeddings.