Every vendor claims around 99% accuracy on their own data. An independent lab reached a different conclusion: 17 systems, one protocol, private labels. Here is the ranking, and how to read it.
According to the independent Podonos benchmark (June 2026), four AI voice detection systems clear 95% accuracy on a private test set of 4,524 clips. In the lead, a US system, Resemble AI. Right behind, by just 0.35 of a point, Whispeak: 2nd of 17, with the most balanced error profile of the entire field, and the highest European ranking on the leaderboard.
Why advertised "99% accuracy" figures mean nothing
Most vendors communicate around 99% accuracy. The problem: these figures are measured by the vendor itself, on its own test set. They are not comparable from one provider to another, and they are easy to overfit on data the model already knows.
That is the whole point of an independent benchmark. Podonos maintains a fixed evaluation set with private gold-standard labels and applies exactly the same scoring to every system. No vendor supplies its own files, and none knows the answers in advance. For once, the results are directly comparable.
What makes this benchmark trustworthy
Not all comparisons are equal. Five criteria separate a credible evaluation from a marketing argument:
Private labels. The test set and the correct answers are not public, which prevents any training on the evaluation data.
One protocol for everyone. Same files, same default thresholds, no per-tool tuning.
Realistic data. Real voices from public corpora (VCTK, LJSpeech, LibriTTS) against clones generated by around 25 recent synthesis systems, including ElevenLabs and F5-TTS.
Multiple audio formats. Six formats tested (mp3, wav, flac, ogg, m4a, webm) to measure robustness in real-world conditions.
Published error modes. The benchmark does not give a single global score, it details how each system fails. And that is precisely what separates a good detector from one that is genuinely usable.
The criterion that really matters: the error profile
A global accuracy score hides the way a detector fails. Two systems with the same accuracy can be radically different in production, because they do not get it wrong in the same way. Two types of error must be separated:
False negative
A voice deepfake goes undetected and passes as a real voice. In a bank or an insurer, that is fraud getting through: impersonation, vishing, CEO fraud.
False positive
A real human voice is wrongly flagged as fake. That is a legitimate customer blocked, friction, support tickets and, at scale, churn.
A detector that almost never gets real voices wrong but lets a significant share of deepfakes through is not "safer", it simply shifts the risk. Real performance means keeping both error rates low at the same time. Keep this criterion in mind: it is what separates the top of the leaderboard.
The ranking: four systems clear 95%
In the previous edition, only two systems passed 95% accuracy. There are now four. Here is the top of the leaderboard, with the breakdown of both error types:
| System | Accuracy | False positives | False negatives |
|---|---|---|---|
| Resemble AI (US) | 98.05% | 2.5% | 1.4% |
| Whispeak | 97.70% | 2.9% | 1.7% |
| Aurigin AI (CH) | 96.75% | 1.5% | 5.0% |
| Pindrop (US) | 95.05% | 6.2% | 3.7% |
Source: Podonos benchmark, June 2026. Private test set of 4,524 clips, balanced real voices / synthetic clones.
Resemble AI is first with 98.05%. Right behind, Whispeak, by just 0.35 of a point: the gap at the top is razor-thin. But the number that matters most lies elsewhere. With 2.9% false positives and 1.7% false negatives, Whispeak posts the most balanced error profile of the entire field. On the criterion that truly separates a usable detector, our solutions are at the top: they do not over-alert and they do not let fraud through, exactly what you expect from a system running continuously on real calls.
The contrast is clear: against American and often multimodal players, Whispeak is a European team focused on one thing only, the voice. That focus shows directly in the results.
PS. The latency figures reported for Whispeak in the full benchmark were measured on a development environment and do not reflect our production performance. We would be glad to share the real numbers.
Open source still does not generalise
The benchmark includes nine open-source models, some trained on recent data (ASVspoof 5, VoxCelebSpoof). All score between 47.6% and 62.9% accuracy, barely better than chance for a binary decision. Several even collapse onto a single class: they predict almost everything as real, or almost everything as fake.
The lesson is clear, and it holds for any buyer: an academic checkpoint taken as-is is no substitute for a production detector against today's voice clones. More recent training data alone is not enough to close the gap.
What a buyer should really ask
Before comparing accuracy scores, four questions separate a sound choice from a gamble:
1. Is the evaluation independent? A figure self-reported by the vendor is not comparable. Look for a third-party test, on private labels.
2. Which error can you least afford? Letting fraud through, or blocking a legitimate customer? The answer depends on your business context and drives the choice far more than the global score. This is where a balanced error profile makes the difference.
3. Does the system hold up on realistic data? Multiple formats, multiple voice generators, conditions close to production.
4. Does the provider meet your sovereignty requirements? Hosting, data processing, GDPR, DORA and NIS2 compliance: for a European organisation, this criterion weighs as much as raw performance.
About Whispeak. Whispeak is a European deeptech specialising in voice security: voice deepfake detection and voice biometric authentication, both powered by the same acoustic engine. Beyond the Podonos benchmark, our solutions rank in the top 4 of the ASVspoof Interspeech 2024 challenge (4.16% EER), won the DGA Cyber Challenge 2024, and are ISO/IEC 27001:2022 certified.
Our detection solution does not need to store voice data to operate, a point that matters for European buyers subject to GDPR, DORA or NIS2.
Frequently asked questions
What is the best AI voice detector?
According to the independent Podonos benchmark of June 2026, Resemble AI is the most accurate (98.05%), followed by Whispeak (97.70%). Whispeak, however, posts the most balanced error profile of the field and is the highest-ranked European detector. Four systems in total clear 95% accuracy.
What is the Podonos benchmark?
It is an independent evaluation of voice deepfake detection systems. Podonos maintains a private test set of 4,524 clips, balanced between real voices and synthetic clones, and applies the same protocol to every system. In the June 2026 edition, 17 systems were compared.
Why does the error profile matter more than accuracy?
Because two systems with equal accuracy can fail differently. A false negative lets fraud through, a false positive blocks a legitimate customer. In production, keeping both rates low at the same time matters more than the global score. Whispeak posts the most balanced profile in the benchmark.
Are open-source models reliable for detecting voice deepfakes?
According to the Podonos benchmark, no. The nine open-source models tested score between 47.6% and 62.9% accuracy, barely better than chance, even when trained on recent data. They are not a replacement for a production detector.
Evaluating an AI voice detection solution for your organisation?
Talk to an expertSource: Podonos, "Audio Deepfake Detection Benchmark", June 2026. Full benchmark, error-profile charts, dataset and reproduction code available on the Podonos website.
