A Multi-task Conformer for Spoofing Aware Speaker Verification

Spoofing attacks conducted via logical methods such as voice synthesis and voice conversion could significantly degrade the performance of a speaker verification system. Many spoofing countermeasures have been proposed, but they are only trained separately from speaker verification systems. There are only a few efforts to design a single model capable of rejecting both utterances spoken by different speakers as well as spoofing utterances. Meanwhile, Conformer, a combination of Transformer and Convolution Neural Network, has shown remarkable success for automatic speech recognition, but its application in speaker verification and anti-spoofing systems has not yet been explored. In this work, we proposed a multitask Conformer with statistical pooling for both speaker verification and voice spoofing detection. Our system achieved a 70% relative SASV-EER improvement over baselines on the ASVspoof 2019 LA evaluation set.