- Home
- AI Deepfake Detection Tools
- Pella Research
Pella Research
Amsterdam, Netherlands
What it is
Pella Research develops models that assess whether audio contains synthetic speech. Its research examines vocoder artifacts, spectral statistics, and prosodic cues, while its evaluation sets track newer voice-cloning systems with documented provenance and no train/test leakage. A browser demo accepts a recorded or uploaded voice clip and returns a verdict without signup; the site says audio is not stored. For programmatic use, its REST API returns calibrated authenticity scores and per-segment analysis. Access is reviewed after a request, typically within 24 hours, and accredited journalists and academics can receive a free tier. Pella says its models are trained and tested with phone-line compression, background noise, and re-encoding. It reports pellav2 results of 95.82% accuracy, a 0.959 F1 score, and 57 ms latency in the August 2026 Podonos benchmark, plus 93.6% accuracy on m4a files with errors leaning toward false positives. The model is listed as MIT-licensed open weights.
For and against
For
- Browser demo needs no signup
- API returns calibrated authenticity scores
- Per-segment analysis supports detailed review
- Models address compression, noise, and re-encoding
- pellav2 is listed as MIT-licensed open weights
Against
- API access requires review
- Demo wake-up can take up to a minute
- Reported m4a accuracy is 93.6%
- Errors reportedly lean toward false positives
The verdict
Pella Research offers a no-signup demo, reviewed API access, and published evaluation material for synthetic-speech checks. Readers should account for its reported m4a performance and false-positive tendency.
Who it suits
Pella suits journalists, platforms, courts, researchers, and institutions investigating whether audio is synthetic. Its API is available to approved users, with a free tier for accredited journalists and academics.
On the record
- Access and support
- API access is reviewed on request for journalists, platforms, researchers, and institutions, with a typical response within 24 hours.pellaresearch.com · read 7 Oct 2026
- Access response
- Pella says it reviews API requests and typically responds within 24 hours.pellaresearch.com · read 7 Oct 2026
- API
- The detection API is a REST endpoint that returns calibrated authenticity scores and per-segment analysis, with access granted on request.pellaresearch.com · read 7 Oct 2026
- API audience
- The API is offered to journalists, platforms, researchers, and institutions, and the page lists a free tier for accredited journalists and academics.pellaresearch.com · read 7 Oct 2026
- Benchmark limitation
- Pella reports 93.6% accuracy on m4a files and says its errors lean toward false positives.pellaresearch.com · read 7 Oct 2026
- Benchmark result
- Pella reports that pellav2 achieved 95.82% accuracy, a 0.959 F1 score, and 57 ms latency in the August 2026 Podonos benchmark.pellaresearch.com · read 7 Oct 2026
- Benchmarks
- Pella says its evaluation sets track new voice cloning systems and use documented provenance with no train-test leakage.pellaresearch.com · read 7 Oct 2026
- Demo limitation
- The live demo runs on free shared hardware, and its first request after a quiet period can take up to a minute to wake up.pellaresearch.com · read 7 Oct 2026
- Free access
- Pella states that accredited journalists and academics can receive a free API tier.pellaresearch.com · read 7 Oct 2026
- Intended users
- Pella describes its work as serving platforms, journalists, courts, and researchers investigating whether audio is synthetic.pellaresearch.com · read 7 Oct 2026
- Known failure case
- Pella reports 93.6% accuracy on m4a files in that benchmark and says its errors lean toward false positives.pellaresearch.com · read 7 Oct 2026
- License
- The maker lists pellav2 as open weights under the MIT license.pellaresearch.com · read 7 Oct 2026
- Live demo
- The browser demo lets users upload or record a voice clip and receive a verdict without signup; the page says no audio is stored.pellaresearch.com · read 7 Oct 2026
- Location
- The site lists Amsterdam, Netherlands.pellaresearch.com · read 7 Oct 2026
- Model availability
- The site describes pellav2 as open weights under the MIT license and says the public checkpoint reproduces the submitted benchmark score.pellaresearch.com · read 7 Oct 2026
- Product
- Pella Research builds models to detect synthetic speech by analyzing vocoder artifacts, spectral statistics, and prosodic cues.pellaresearch.com · read 7 Oct 2026
- Purpose
- Pella Research studies synthetic speech and builds models to detect it in real-world audio.pellaresearch.com · read 7 Oct 2026
- Real-world conditions
- Pella says its models are trained and tested with phone-line compression, background noise, and re-encoding.pellaresearch.com · read 7 Oct 2026
