Trust and safety
Machine-readable marking
How Pynokio marks AI-generated content in a form that software can check, and how a file can be verified. The same information is published as JSON for machines.
/.well-known/ai-content-provenance.json ยท version 2026-09-26
Audio watermark
Every generated speech file carries an inaudible watermark embedded in the sound itself, with the Pynokio identifier.
| Technology | AudioSeal (open source) |
|---|---|
| Generator / detector model | audioseal_wm_streaming / audioseal_detector_streaming |
| Sample rate | 16 kHz |
| Identifier | 16-bit, 0x5643 |
| Applies to | Text to speech, voice clone previews, voice design previews |
| Decision rule | Present when at least 15 of 16 identifier bits are read with clear confidence |
Platform provenance record
| HTTP header | X-Pynokio-Synthetic-Media: 1 on file downloads |
|---|---|
| Record link | HTTP Link header of file downloads; provenanceUrl in API asset objects |
Signed file metadata (C2PA)
Embedding a signed C2PA manifest in downloaded audio files is planned. Until it is available, files do not contain it.
Detection
Checks are free. Anyone can upload a file to the Audio Detector (WAV, MP3, OGG, FLAC, M4A or WEBM, up to 32 MB). The result says whether the watermark is present, possibly present or not confirmed, and shows how much of the Pynokio identifier was read.
Limits: heavy processing (telephone-band filtering, loud noise, noise reduction, speed or pitch changes) can make the watermark unreadable, so a missing watermark does not prove that a file was not made with Pynokio. Regulators, researchers, media and fact-checkers who need higher volumes can contact support.