Every AI Sound on Tunio Now Carries an Inaudible Signature

Every AI Sound on Tunio Now Carries an Inaudible Signature

A lot of what plays on a Tunio station is made by machines. AI composes music and jingles, a virtual host reads the weather, news bulletins are scripted and voiced automatically, whole podcast episodes are generated end to end. That has always been the point of the platform: radio that runs itself.

But when synthetic audio becomes indistinguishable from the real thing, a fair question follows: how does anyone — a listener, a regulator, a rights society — tell what was made by AI? In Europe that question now has legal weight. So we answered it at the infrastructure level: every piece of AI-generated audio Tunio produces now carries an inaudible, machine-readable watermark, embedded before it ever reaches your stream.

You will never hear it. But any tool that knows how to look can verify, from a few seconds of recorded air, that a sound was generated on Tunio — and even which track it was.

Why now: the EU AI Act's Article 50

The EU Artificial Intelligence Act is the world's first comprehensive AI law, and its transparency chapter — Article 50 — applies from 2 August 2026. Paragraph 2 is the one that matters for anyone generating audio at scale:

Providers of AI systems generating synthetic audio, image, video or text content shall ensure that the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated.

In plain terms: if your platform makes AI audio, that audio must carry a mark that software can find, and the marking must be effective, robust and reliable. Systems already on the market get a short grace period, until 2 December 2026, to comply. Penalties for ignoring the transparency rules reach €15 million or 3% of worldwide turnover.

For Tunio users this is the good kind of regulation news: the marking obligation sits with the provider — us — and it's already done. Music, announcements and bulletins generated on the platform go out with the machine-readable marking Article 50 asks for, with nothing to configure and nothing added to your workflow.

What exactly gets marked

Every generation pipeline on the platform now embeds the watermark in its final output:

• AI music — every generated track, including the existing AI music catalog, which we re-marked retroactively
• Jingles and stingers — station IDs, sweepers, branded intros
• Voice announcements — everything produced by the TTS pipeline, in every voice and language
• News blocks — automatically scripted and voiced bulletins
• Weather forecasts — the automated bulletins read by the virtual host
• Generated podcasts — AI-scripted, AI-voiced episodes

And just as deliberately, one thing is never marked: your own uploads. Music you license, shows you record, ads your clients bring — human-made content stays untouched, because falsely labeling real audio as synthetic would be exactly the kind of confusion the law is trying to prevent.

How do you watermark sound, anyway?

The technique is called spread-spectrum audio watermarking. Instead of adding a tone or a code you could hear, the encoder makes thousands of tiny adjustments across the frequency spectrum — each one far below the threshold of human hearing, all of them together forming a repeating digital pattern. A decoder that knows the secret key can read the pattern back out of the audio; an ear never will.

Into that pattern we write a 128-bit payload: the unique ID of the generated track. So the mark doesn't just say "this is AI" — it says which AI output it is, tying the audio on the air back to a specific item in the catalog.

Getting this to survive real broadcasting was the hard part, and it shaped the whole design:

• The mark is embedded at the last step. Radio production is brutal to hidden signals: mixing a voice over a music bed, time-stretching to fit a slot, crossfading — each of these destroys a watermark applied too early. We measured it. So the signature goes into the finished artifact, after all mixing and tempo processing, right before encoding.
• It survives the stream. Your listeners hear lossy-compressed audio, not studio files. The watermark is strong enough to be read back from the broadcast stream after transcoding — from a snippet recorded off the air, not just from the original file.
• It's keyed. Embedding and detection use a secret cryptographic key, so the mark can't be trivially forged or stripped by third parties.
• It never blocks your radio. Watermarking is a best-effort step with strict time limits. If anything about it ever fails, generation simply continues — your station's reliability always wins.

Detection: closing the loop

A mark nobody can check is just decoration, so we built the other half too. Our API can take a short recorded snippet — a phone capture of a stream, an exported clip — and answer in under a second whether it carries a Tunio signature, resolving the embedded ID against the audio catalog to identify the exact track.

That's what Article 50 means by "detectable", and it's useful well beyond compliance: proving a piece of audio came from your station, tracing where generated content ends up, or confirming that what a rights society flagged is in fact AI-composed music rather than someone's copyrighted recording.

Transparency as a feature

Regulation set the deadline, but honestly, this is where broadcasting was headed anyway. Trust in audio is becoming a product feature: advertisers ask what they're placed next to, platforms ask what they're hosting, listeners increasingly want to know when a voice is human. An inaudible signature in every generated sound is the cleanest possible answer — it tells the truth about the content without interrupting it.

Tunio stations get that answer built in. Create your radio, let the AI do the heavy lifting, and stay effortlessly on the right side of the strictest AI transparency rules in the world.