On Joint Optimization of Automatic Speaker Verification and Anti-Spoofing in the Embedding Space

Biometric systems are exposed to spoofing attacks which may compromise their security, and voice biometrics based on automatic speaker verification (ASV), is no exception. To increase the robustness against such attacks, anti-spoofing have been proposed for detection of replay, synthesis conversion-based attacks. However, most techniques loosely integrated with ASV system. In this work, we develop a new integration neural network jointly processes embeddings extracted from in order detect both zero-effort impostors Moreover, propose loss function minimization area under expected (AUE) performance spoofability curve (EPSC), allows us optimize desired operating range biometric system work. evaluate our proposals, experiments were carried out recent ASVspoof 2019 corpus, including logical access (LA) physical (PA) scenarios. The experimental results show that proposal clearly outperforms some well-known at score- embedding-level. Specifically, achieves up 23.62% 22.03% relative equal error rate (EER) improvement over best performing baseline LA PA scenarios, respectively, as well gains 27.62% 29.15% AUE metric.