Target Speaker Verification With Selective Auditory Attention for Single and Multi-Talker Speech

Speaker verification has been studied mostly under the single-talker condition. It is adversely affected in presence of interference speakers. Inspired by study on target speaker extraction, e.g., SpEx, we propose a unified framework for both single- and multi-talker speech, that able to pay selective auditory attention speaker. This (tSV) jointly optimizes module representation via multi-task learning. We four different embedding schemes tSV framework. The experimental results show all significantly outperform other competitive solutions speech. Notably, best scheme achieves 76.0% 55.3% relative improvements over baseline system WSJ0-2mix-extr Libri2Mix corpora terms equal-error-rate 2-talker while performance speech par with traditional system, trained evaluated same