The result
Podonos tested 23 systems for detecting AI-generated voices. Fennura ranks 3rd, with 98.63% accuracy, 1.2% false alarms and 1.5% missed fakes.
In the Podonos benchmark, that makes Fennura number 1 in Europe for detecting audio deepfakes.
Accuracy in the Podonos Audio Deepfake Detection Benchmark, September 2026.
Podonos writes that the top of the field is close: less than one percentage point separates the first three systems, and all three keep both error rates under 1.6%.
Detection on-device
One remark in the report matters more to us than the ranking. Podonos notes that Fennura was submitted as a CPU-only detector that runs on the device and needs no graphics card for the analysis.
For companies, that is often the deciding point. Anyone who discusses client, health or contract data on the phone often may not pass those conversations on externally. A detection system that sends audio to the cloud then fails not on the technology, but on data protection, a works agreement or a duty of confidentiality.
With Fennura, the conversation stays where it takes place. No data leaves the device. That also lowers the hardware requirements: an ordinary office computer is enough.
Why this test is meaningful
Vendors of detection software often publish their own figures. The problem with that is obvious: whoever assembles the test also determines the result.
The Podonos benchmark works differently. The correct answers stay with the operator. No vendor knows in advance which file is real and which is not, and nobody can tune their system to the test. 4,524 files were tested, half real and half synthetic, in six file formats. The synthetic recordings come from around 25 commercial voice-cloning tools, that is, from possible tools that attackers could actually use today.
Fennura evaluated every one of these file types. Some systems in the field rejected part of the files and were therefore scored only on a smaller selection.
Two kinds of error, not one
With detection software, usually only the hit rate is quoted. On its own it says little: a system that flags every call as a fake detects every deepfake and is still useless.
Both directions matter. False alarms cost time and trust, missed fakes cost money. At 1.2% to 1.5%, both of Fennura's figures sit close together.
What the benchmark leaves open
A test result is a snapshot. Finished files were tested, not live operation with background noise and changing speakers, and the test was run against today's voice-cloning models. Our analysis takes 553 milliseconds on average and is therefore slower than that of some systems in the field. Even so, that results in an RTF (real-time factor) of 0.14 and thus real-time capability on an ordinary office laptop.
We consider these points worth stating. A security promise is worth only as much as the openness about its limits.
Would you like to talk this through with us, or use Fennura in your own organization? Write to us at hello@fennura.ai.
All benchmark results are publicly available: https://github.com/podonos/audio-dfd-benchmark
Benchmark data as of September 2026. Source: Podonos, Audio Deepfake Detection Benchmark.