This is the authors' abstract. We don't add key points for this paper.
Arabic, spoken by more than 400 million people, remains critically underrepresented in speech deepfake detection research. No existing Arabic benchmark provides unified evaluation across both fully synthetic spoofing detection and frame-level forgery localization. This paper introduces NAUSS-Spoof50, a curated multi-task Arabic speech benchmark comprising 76 730 utterances from 50 native Arabic speakers representing 13 national backgrounds across four broad regional groups, recorded across three devices and two acoustic environments under a strict speaker-disjoint protocol. Six spoof-generation methods span four paradigms: neural vocoding, neural codec reconstruction, text-to-speech, and voice cloning. A partially manipulated subset provides frame-accurate ground-truth annotations for five signal-level transformations. Three architecturally diverse baselines are evaluated for utterance-level spoof detection: RawNet2, AASIST, and a wav2vec 2.0-based countermeasure. AASIST achieved the lowest point EER of 0.283% [95% CI: 0.177%, 0.400%], followed by wav2vec 2.0 at 0.431% [95% CI: 0.301%, 0.584%]; their paired difference was not statistically significant. Cross-dataset evaluation revealed substantial performance degradation, indicating limited robustness under corpus and language shift. XTTS v2 and ElevenLabs produced the highest numerical EERs across the evaluated generation methods. Condition-stratified analysis showed that clean-environment recordings produced higher point EERs than noisy recordings across all models. For frame-level forgery localization, the wav2vec 2.0-based model achieved a Boundary-F1 of 0.913, with partial deletion producing the lowest localization performance. NAUSS-Spoof50 establishes a structured benchmark for Arabic speech anti-spoofing and forgery-localization research under diverse attack and acquisition conditions.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Moallim et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: