Resemble AI has published benchmark results for DETECT-3B Omni, the company’s three-billion-parameter multimodal deepfake detection model, with audio results compared against named commercial systems including Aurigin AI, Hive, and Reality Defender, and against open-source baselines. The numbers are Resemble’s own; the company notes that it does not control the composition of the test sets it reports against.
The audio numbers map most directly onto voice anti-spoofing, the category that voice biometric systems and contact center fraud teams rely on to flag synthetic speech. On the Podonos Audio Deepfake Detection Bench, which covers more than 40 languages and includes telephony codecs such as G.711 and G.723.1 across roughly 4,500 files, Resemble reports 98.1 percent pooled detection accuracy for DETECT-3B Omni and a 1.4 percent rate of misclassifying generated audio as authentic. On a second audio benchmark, Speech DF Arena, the company reports 97.4 percent average accuracy and an equal error rate of about 2.1 percent across languages.
The named competitor numbers Resemble published on the same audio benchmarks place Aurigin AI at 96.8 percent, Hive at 83.5 percent, and Reality Defender at 71.3 percent, with open-source models Wav2Vec2 and AASIST scoring lower. As with any vendor-run benchmark, these numbers reflect the specific test sets, file formats, and evaluation methodology Resemble chose, and the company itself notes that detection in its system relies on generation-artifact patterns rather than language-specific features.
The other modalities round out the omni claim. Resemble reports 85.1 percent overall accuracy on the DFBench Image 2025 image deepfake benchmark and about a 2.5 percent equal error rate on video generation models, again drawn from the company’s own evaluations.
The underlying question is whether a single model can replace the patchwork of language- or codec-specific detectors that have historically been needed to cover real-world telephony. The Podonos suite’s inclusion of telephony codecs and multiple file formats is the closest part of Resemble’s benchmark to that operational setting, and a 98.1 percent pooled accuracy is the company’s headline claim in that lane.
Independent third-party evaluation, comparable to what NIST’s voice biometric programs provide for speaker recognition, is not yet a standard fixture of deepfake detection.
Sources: Resemble AI
–
By the ID Tech Editorial Team
