deetech, the deepfake detection venture we built at [r]think, is ranked number one of 23 systems on the Podonos audio deepfake detection benchmark. It scored 99.56% accuracy, flagging 0.4% of real voices and missing 0.4% of fakes.

For a venture studio, a result like this is more than a trophy. It is evidence that a small, AI native team can build deep technology that holds its own on a scoreboard someone else controls.

The competition is formidable. Some of deetech's competitors are unicorns, and together they have raised more than US$600 million. We are happy to see our small team in the lead.

The result

  • 99.56% accuracy across 4,524 real and fake voice recordings.
  • 0.4% of real voices flagged as fake. That is the false positive rate.
  • 0.4% of fakes missed. That is the false negative rate.
  • 50 milliseconds per recording on average, self reported, as the board notes for vendor entries.

Why an independent benchmark matters

Anyone can publish a great number on their own test set. What makes this one count is who holds the answers. Podonos runs the benchmark as a neutral evaluation on a fixed set of real and fake recordings. Vendors run their systems and submit their verdicts, and only Podonos holds the labels, so no one can tune to the answer key. Podonos computes the scores and publishes the full ranking openly.

The real win is generalisation

Deepfake detectors share a known weakness. They learn the fingerprints of the generators they were trained on, then stumble when a new one arrives. New voice models ship every month, and fraudsters use whatever is newest.

deetech's new model architecture is a big step on exactly that problem. It is built to generalise: to stay accurate on new deepfakes and spoofing techniques, not only the ones it has already seen. Podonos adds outside proof: first place among 23 systems on a test deetech did not build, scored against labels it never saw.

Fast enough for live calls

At 50 milliseconds per recording, detection keeps pace with a live conversation. That opens a market recorded media forensics never could: screening calls as they happen, in contact centres, on help desks and in video meetings. It is the foundation for BotBlock, deetech's call screening product.

What it says about building ventures

We build ventures where AI creates a problem that existing products were not designed for. Deepfakes are that problem for anyone who trusts a voice, a photo or a document. The playbook is one we keep coming back to, as we argued in Reliability Over Intelligence: go deep on one hard technical problem, test against data you do not control, and let an outside scoreboard decide whether you are ready.

Open to independent testing

Public benchmarks keep the industry honest, and deetech wants more of them. The team is open to any external, impartial benchmark across audio, images, video and documents, including tests where the evaluator runs deetech on data it has never seen. If you run one, get in touch. The full write up is on the deetech blog.

Anyone can top their own leaderboard. The ones worth winning are run by someone else.