Open specification

How we mark AI content

Tõlk.fm marks everything its AI produces so people and machines can tell. The format is open: anyone can build a detector or verifier.

Audio watermark: TWM1 is a spread-spectrum watermark added to every AI voice we produce, live and exported. Detection gives a score against a fixed threshold; unmarked audio stays well below it.
Signed provenance: Each marked file carries a compact JWS whose claim follows https://tolk.fm/ai-provenance/v1: whether the content is AI-generated, what kind it is, its languages, whether a cloned voice was used and a hash binding the claim to the content. It holds no personal data. Keys are published at /.well-known/tolk-ai-provenance-keys.json.
C2PA Content Credentials: Where the format supports it, files also carry a C2PA manifest with the IPTC digital source type for AI-generated media.
Export list and notices: Export ZIPs include a signed AI-PROVENANCE.json listing every file, and AI-NOTICE.txt. Cloned-voice audio gets an audible AI notice at the start and every 15 minutes.
Live disclosure: Live listeners see an “AI-generated voice” label and an “AI translation” tag on captions, and the live streams announce the same facts to apps in machine-readable form.