Pella Research develops tools for assessing whether audio is synthetic, with work focused on signals such as vocoder artifacts, spectral statistics, and prosodic cues. Its pellav2 model is available as open weights under the MIT license, and the maker says the public checkpoint reproduces its submitted benchmark score. In the August 2026 Podonos benchmark, Pella reports 95.82% accuracy, a 0.959 F1 score, and 57 ms latency. It also reports 93.6% accuracy on m4a files and says false positives make up more of its errors. Pella says its models are trained and tested on audio affected by phone-line compression, background noise, and re-encoding. A browser demo lets visitors upload or record a voice clip and receive a verdict without signing up; the page says audio is not stored. The demo uses free shared hardware, and its first request after a quiet period can take up to a minute to wake. For programmatic checks, Pella offers a REST API that returns calibrated authenticity scores and per-segment analysis. API access is reviewed on request, typically with a response within 24 hours. Accredited journalists and academics can receive a free API tier. Pella also publishes evaluation sets tracking new voice-cloning systems, with documented provenance and no train/test leakage. The company lists Amsterdam, Netherlands, as its location.
Who it is for
Pella suits journalists, platforms, researchers, institutions, and courts investigating whether audio is synthetic. Its browser demo supports individual checks, while its reviewed API provides scores and segment-level analysis for programmatic use.
What is good
- Browser demo accepts uploaded or recorded clips without signup.
- Demo page says audio is not stored.
- API returns calibrated authenticity scores and per-segment analysis.
- pellav2 is open weights under the MIT license.
- Evaluation sets document provenance and avoid train/test leakage.
What to know first
- API access requires a request and review.
- Demo may take up to a minute to wake after a quiet period.
- Pella reports 93.6% accuracy on m4a files and a false-positive tendency.
Verdict
Choose Pella Research if you need synthetic-speech analysis, an open-weight model, or an API for audio checks. The browser demo offers a direct way to get a verdict, while API users should account for reviewed access and the reported m4a accuracy and false-positive limitation.
Get started with Pella Research
- Open the browser demo on the Pella Research website.
- Upload a voice clip or record one in the browser demo.
- For API access, submit a request for review.
- If accredited as a journalist or academic, request the free API tier.
- Use the public pellav2 checkpoint under its MIT license.
Limits to know first
The API's free tier is listed for accredited journalists and academics. Pella reports 93.6% accuracy on m4a files and says errors lean toward false positives; the free shared-hardware demo can take up to a minute to wake after a quiet period.
Questions about Pella Research
What does Pella Research do?
It studies synthetic speech and builds models to assess whether audio is synthetic, using cues such as vocoder artifacts, spectral statistics, and prosody.
Is there a free way to try it?
The browser demo is available without signup. Pella also lists a free API tier for accredited journalists and academics.
How do I get API access?
Submit a request for review. Pella says it typically responds within 24 hours.
What does the API return?
It is a REST endpoint that returns calibrated authenticity scores and per-segment analysis.
Is pellav2 open source?
Pella describes pellav2 as open weights under the MIT license and says the public checkpoint reproduces its submitted benchmark score.
What platforms are available?
Pella lists web and API access.
Compared on AI deepfake detection tools
- Free plan
- Yes
- Media types
- audio
- API access
- Yes
- Face analysis
- No
