HYU Submission for the SASV Challenge 2022: Reforming Speaker Embeddings with Spoofing-Aware Conditioning

Abstract

In this paper, we introduce the spoofing-aware speaker verification (SASV) system submitted by the Hanyang University team for SASV Challenge 2022. Our strategy is to learn spoofing-aware speaker embeddings (SASEs) that can effectively produce SASV scores by using a simple cosine similarity scoring backend. To achieve this, we develop a neural-network-based SASE model that uses a spoofing countermeasure (CM) embedding and speaker embedding to produce an SASE. The baseline anti-spoofing model is used to extract CM embeddings, and ResNet-34- and Res2Net-based models are employed to extract speaker embeddings. When evaluated on the ASVspoof2019 logical access dataset, our best proposed SASV system achieved SASV equal error rates of 0.1817% and 0.2793% on the development and evaluation set partitions, respectively, placing 3rd in the SASV Challenge 2022.