Raid AI places first on ArA-DF 2026 for Arabic deepfake audio
The first shared task on Arabic audio deepfake detection scored systems on dialects and speakers they never saw in training. Raid AI's audio model ranked first on that track.
ArA-DF 2026 is the first shared task on Arabic audio deepfake detection. Its Track 1, Dialect Generalization, scores a system on Arabic speech from dialects and speakers that were not in its training data. In the evaluation phase, run on Codabench in July 2026, Raid AI's audio detection model ranked first.
What the task measured
Each audio clip in the test set is either spoofed, meaning synthetic speech from a text to speech or voice conversion system, or bonafide, meaning genuine speech from a native Arabic speaker. A system submits a score for every clip, higher meaning more likely genuine, and the platform computes the equal error rate: the point at which the rate of genuine clips wrongly rejected equals the rate of synthetic clips wrongly accepted. Lower is better, and the ranking is by that number alone.
The task ran in two phases. A development phase gave public feedback on ten percent of the test set. The evaluation phase, from 20 to 25 July 2026, scored the remaining ninety percent and set the final ranking. The organisers, reachable through DFKI, the German Research Center for Artificial Intelligence, approved every team by hand.
Why dialect generalisation is the hard part
Most published audio deepfake detectors are trained almost entirely on English, and on studio quality English at that. They learn what synthetic speech sounds like in one language, and when the language changes they have to fall back on whatever survived the shift. Arabic makes the shift harder than most. Its dialects differ in phonology, rhythm and vocabulary far more than the varieties of English do, and a Gulf speaker, an Egyptian speaker and a Maghrebi speaker will exercise a model in three different ways.
A detector that has only heard Modern Standard Arabic is in a similar position to one that has only heard English: the dialect is a new distribution, and a text to speech system that has been trained on that dialect produces artefacts the detector has never had to separate from genuine speech. Track 1 is built to expose exactly that. The dialects and the speakers in the test set are held out, so a system cannot lean on having heard them before. It has to have learned something about synthetic speech that holds across dialects.
What we entered
We entered with our in house audio detection model, the one behind the platform's voice and audio detection. The model is trained on the acoustic signal rather than on the content of the speech, scoring the frequency distribution and vocal timing that synthetic speech carries and that survive compression and a poor line. That is the same property that lets it work on a real phone call, and it turned out to be the property that matters when the language changes underneath it.
Arabic has been a design goal of the model from the start, not a language added afterwards. The site has said for some time that our accuracy on Arabic dialects is something English first models do not have. This is the first public benchmark, run by an independent organiser on a held out set, where that claim was tested against other systems.
What the result means, and what it does not
It means that on an independent test set of unseen Arabic dialects and speakers, our model separated synthetic from genuine speech better than every other system entered. That is a real result, on a named dataset, with a date, and we are glad to have it.
It does not mean that any number from the benchmark transfers to your traffic. A shared task's test set is a fixed collection of clips; your calls arrive at whatever quality your telephony gives them, in whichever dialect your customers speak, with whichever generators an attacker chose this month. We say this on every page of this site because it is true of every vendor's benchmark, ours included. The way to know how a detector performs on your audio is to run it on your audio.
Test it on your own Arabic audio
Bring a recording you already know the answer to, in the dialect your customers speak, at the quality of a real call. That is the only evaluation that settles anything, and it fits in a thirty minute demo.
What comes next
The shared task has a second track and, we expect, a second year. We will enter both. The model that placed first is the model that runs in the product today, on live calls, uploaded recordings and voice notes forwarded to the WhatsApp bot, and it is evaluated in shadow deployment against real traffic before any change reaches a customer verdict.
