Jointing Multi-task Learning and Gradient Reversal Layer for Far-Field Speaker Verification

Far-field speaker verification is challenging, because of interferences caused by different distances between the speaker and the recorder. In this paper, a distance discriminator, which determines whether two utterances are recorded at the same distance, is used as an auxiliary task to learn distance discrepancy information. There are two identical auxiliary tasks, one is added before the speaker embedding layer to learn distance discrepancy information via multi-task learning, and then the other is added after that layer to suppress the learned discrepancy via a gradient reversal layer. In addition, to avoid conflicts among the optimization directions of all tasks, the loss weight of every task is updated dynamically during training. Experiments on AISHELL Wake-up show a relatively 7% and 10.3% reduction of equal error rate (EER) on far-far speaker verification and near-far speaker verification respectively, compared with the single-task model, demonstrating the effectiveness of the proposed method.