Revealing Cross-Lingual Bias in Synthetic Speech Detection under Controlled Conditions
Victor Moreno, João Lima, Flávio Simões, Ricardo Violato, Mário Uliani Neto, Fernando Runstein, Paula Costa
Speech-based biometric systems have been increasingly deployed in high-stakes domains such as banking, forensics, and authentication. However, these systems remain vulnerable to synthetic speech attacks, such as spoofing and deepfakes. Recent research has focused on developing countermeasures (CMs) capable of detecting manipulated audio. In this work, we investigate whether language identity influences the detectability of synthetic speech in a state-of-the-art CM pipeline. We train a detector on English-only data and evaluate it under controlled conditions using spoofed samples in ten languages synthesized by a standardized text-to-speech system. Despite uniform synthesis settings, we observe significant language-dependent disparities in detection performance. These results suggest that language identity acts as a latent bias factor, challenging the cross-lingual generalization of current CM systems and underscoring the need for fairness-aware multilingual evaluation protocols.