Resemble AI's deepfake audio detector holds up under a fairness stress test — and the researchers have the numbers to prove it.
A team tested DETECT-3B-Omni using 10,240 audio samples generated by 8 different AI voice-cloning systems, drawing on speakers from 30 US states. They ran equivalence tests across four variables: what the audio says (benign versus flagged content), speaker gender, speaker age, and regional accent. The result — accuracy differences of at most 2 percentage points at 99% confidence — suggests the model is keying on acoustic artifacts rather than who is speaking or what they are saying.
That distinction matters a lot for real-world deployment. A detector that quietly underperforms on certain demographics or on politically charged speech would be both a liability and a GDPR problem — it would effectively discriminate by proxy. An equivalent-accuracy result across these axes is the minimum bar regulators and enterprise buyers should demand before trusting any such tool.
Worth noting: the study covers US English only, and eight voice-cloning systems is a reasonable but not exhaustive sample — newer or more obscure generators may behave differently. Independent replication, rather than a vendor-adjacent audit, would settle the question more cleanly.