Raid AI is now on the Podonos benchmark, with 98.54% accuracy
Podonos is an independent benchmark for voice deepfake detection. Raid AI's model scored 98.54% accuracy on its 4,524 file test set and answered every clip.

Raid AI's deepfake detection model is now on the Podonos benchmark, with 98.54% accuracy. Podonos is an independent benchmark for voice deepfake detection: it keeps a fixed test set, holds the labels privately, and scores every submission itself, the same way and on the same files. Our model joined the leaderboard in its October 2026 update.
The results
Raid AI's audio deepfake detector, scored on all 4,524 files in the benchmark's test set:
- —Accuracy: 98.54%, seventh of the 28 systems on the leaderboard.
- —F1 score: 0.985.
- —False positive rate: 1.9%, the share of real clips flagged as fake.
- —False negative rate: 1.0%, the share of fake clips missed.
- —Rejection rate: 0.0%. The model returned a verdict on every file.
- —Latency: 93 ms per file on average, measured by Raid AI on its own hardware.
- —Real time factor: 0.033, so scoring a clip takes about a thirtieth of the clip's length.
The two error rates show which way the model leans. It flags a few more real clips than it misses fakes, 1.9% against 1.0%, so when it is wrong it is more often wrong on the side of caution. Thirteen of the 28 systems on the leaderboard cleared 95% accuracy, and Raid AI's model is one of them.
How the benchmark works
The test set holds 4,524 audio files, half real and half fake. The real speech comes from three public corpora: VCTK, with 110 English speakers across a range of accents; LJ Speech, a single speaker reading public domain audiobooks; and LibriTTS, with 904 speakers. The fakes come from about 25 text to speech and voice cloning systems, both commercial services and open weights models run locally. Every synthetic clip was transcribed back by a speech recognition model to confirm it says what it was meant to say.
The files come in six formats, mp3, wav, flac, ogg, m4a and webm, so a detector cannot rely on one container or codec. The benchmark is English only.
Podonos keeps the labels. Each vendor runs its system over the files on its own infrastructure and submits one prediction per file, and Podonos computes every metric from those predictions. A vendor can submit one model a month, and the leaderboard is updated monthly.
What the metrics mean
Accuracy is the share of all files classified correctly. F1 balances precision, how many flagged clips were really fake, against recall, how many fakes were caught. Podonos deliberately leaves out equal error rate, because it assumes a decision threshold chosen with the answers already known, which no one can do in production. Every figure above is measured on the verdict the model actually gave for each file.
What the result means, and what it does not
It means that on an English test set we did not build, in six audio formats, with labels we never saw, our model classified 98.54% of 4,524 clips correctly and answered every one of them.
It does not tell you how the model performs on your calls. These are studio quality recordings, and phone lines, other languages and new generators are a different test. Podonos plans telephony tracks, the same clips degraded to narrowband and wideband phone audio, for its November 2026 update. The way to know how a detector performs on your audio is still to run it on your audio.

