ParlaSpoof-BR exposes audio deepfake detector bias in Brazilian politics

Researchers at Universidade Federal de Goiás (UFG) developed ParlaSpoof-BR, a novel audio deepfake dataset sourced from real recordings of the Brazilian Chamber of Deputies. The team expanded it with synthetic utterances generated by diverse text-to-speech (TTS) and voice conversion (VC) models, creating a domain-specific benchmark for Brazilian Portuguese political speech. They then benchmarked state-of-the-art audio deepfake detectors to assess both generalization to this underrepresented language and potential bias in detection outcomes.

The analysis revealed that current detection systems struggle to provide consistent decisions across the dataset's diverse speakers and audio conditions. Critically, the dominant sources of error were methodological factors — the choice of synthesis model and the extent of manipulation — rather than demographic disparities like gender or accent. This suggests prior bias concerns may be overstated, and that improving detector robustness to different generation techniques is the more urgent challenge. ParlaSpoof-BR provides a socially consequential benchmark for studying audio deepfake detection in an electoral context, supporting development of more reliable detection systems for protecting electoral integrity in Brazil.