#1 of 28 on an Independent Benchmark
Podonos runs an independent benchmark for audio deepfake detectors. It keeps the labels for 4,524 English clips, made with about 25 current voice cloning and text-to-speech tools, and scores every vendor itself. In the October 2026 update, both our Arcas models tied for first out of 28 systems with three errors each. Third place, deetech.ai, made about 20 errors on the same clips. Resemble DETECT-World made 24.
The Fewest Errors of Both Kinds
Arcas flagged 0.1% of real voices as fake and, after rounding, missed none of the fakes. No other commercial system was at or below 0.1% on both. For a bank contact center or an IT help desk, that means fewer genuine callers pushed into extra checks and no cloned voice let through.
Fast Enough for Live Calls
Arcas 90M matched the larger model's accuracy at 89 ms per clip, and Arcas 290M took 118 ms. Both run well inside a live call, which is how NeuroCall uses them. As with most systems on the board, these latency figures are reported by the vendor.
Test It on Your Own Audio
A public benchmark gives you a shortlist. Before you buy, run your own test set and price out what each kind of error costs you. We're happy to run Arcas on your audio.
Source record
