Audio has no picture to put an icon on, so the European Code of Practice asks for it to be declared by voice. Here I follow how version 1.0.0-beta1 of AI Disclosure, the module I created on drupal.org within the Drupal AI Initiative, does it. Images have an article of their own, and so does video.

Audio needs to be declared when it is a deep fake: when a voice credibly resembles a person who exists or could exist, and whoever listens would take it as real. The Commission's guidelines give exactly this example: a newspaper podcast with the voices of its hosts and a guest cloned to discuss current events. A voice given to a fictional character, in an audiobook or a game, stays out if nobody is misled about who is speaking. Noise reduction or normalisation, which removes interference without changing the words, stays out too.

The obligation falls on the publisher, under Article 50(4) of the AI Act, and it applies from 2 August 2026. The watermark the voice generator puts in the file is the provider's obligation, and a publisher cannot rely on it for their own declaration.

For audio, the final June 2026 version of the Code of Practice asks for a short spoken notice at the start of the deep fake, in plain words, in the language of the content or in English. In long or live content, where appropriate, reminders at regular intervals are added to the opening notice, tones or earcons included, and at least after every break, for example after an ad. A recognisable sound, an earcon, can support the notice. It can also replace it, but only while waiting for the common signal a European Task Force still has to define, and together with campaigns that explain to the public what it means. When the audio is played on a screen, the icon is needed alongside the spoken notice.

Two rules still going around come from the first draft of 17 December 2025: the opening notice alone for audio under 30 seconds, and the icon for at least five seconds on real-time videos of creative works. The final June 2026 text dropped both.

Whoever opens the page sees the EU icon and the grade sentence above the player, written and always visible. Whoever presses play first hears the notice, for example "This audio was generated by AI.", then a 0.4-second pause, then the recording. The icon stays even when the site serves the copy with the notice, because the person looking at the page has to see the label, not only hear it.

The spoken notice is the part that travels with the file: whoever downloads the audio and republishes it elsewhere carries the declaration along too.

The notice is a clip a few seconds long, one for each active language on the site and for each of two cases: audio generated with AI and audio modified with AI. The module's "Spoken disclaimer" page shows the sentence to say in each language, and offers three ways to provide the clip: upload it, record it in the browser by reading the sentence, or generate it with a speech synthesis service, if the site uses the Drupal AI module with a suitable provider. A generated clip is played back before you keep it, and nothing is generated on its own.

The module rejects a clip longer than 8 seconds, a silent clip, and a file with no audio track. For a file the editor can pick up to three languages, heard one after the other, for example Italian and then English; without a choice, the media's language applies.

In long recordings the notice comes back every 300 seconds by default, except in the last 2 seconds. Two short tones, which the module generates with FFmpeg, can be added before each reminder. The tones sit alongside the notice and never replace it.

The copy contains the notice, the pause and the recording, all converted to 44,100 Hz stereo, and two labels in the file: the grade sentence and the IPTC source type. In MP3 they sit in the ID3 fields, in M4A in the MP4 tags, in FLAC, Ogg and Opus in the Vorbis comments. A WAV keeps only the sentence.

I tried it on a test site with a 15-second MP3. The copy lasted 17.8 seconds: the English notice, the pause and then the whole recording. Both labels were in place, and the uploaded original stayed as it was.

An audio watermark is a signal the ear cannot hear, hidden in the sound. The provider of the system that generates the voice adds it, like the C2PA signature for images, covered in the article on C2PA Sign and Content Credentials with AI Disclosure. A site that generates audio with its own system is also a provider, and for that part it can add a watermark to the copies with audiowmark (opens in a new tab), a free program.

The message written into the watermark is a 128-bit code derived from the site's identifier and the file's, so whoever finds it can trace the source file. On the test site I read it back from the copy, and it matched the one computed separately. With a private key set in the module, only whoever holds it can read or remove it.

It needs FFmpeg. Copies are prepared in a queue, on cron, one at a time, so no visitor waits. A site that cannot run FFmpeg prepares the copy on another machine, with the commands printed by drush ai-disclosure-av:render-command, and hands it back with drush ai-disclosure-av:attach. When the grade changes, the site keeps serving the last ready copy until the new one arrives.

One command is enough to read the labels of a copy:

ffprobe -v error -show_entries format_tags -of default=nw=1 copia.mp3

To hear where the notice ends and the recording begins, the silencedetect filter of FFmpeg finds the pause:

ffmpeg -i copia.mp3 -af silencedetect=noise=-45dB:d=0.2 -f null -

Sources checked on 10 and 11 October 2026, checked again on 20 October 2026.

1. AI Disclosure, documentation: audio (opens in a new tab). Spoken notice, clips, languages, reminders, watermark. Primary source.
2. AI Disclosure, labelled copies of audio and video (opens in a new tab). Queue, small servers, Drush commands. Primary source.
3. Code of Practice on Transparency of AI-Generated Content (opens in a new tab), European Commission, final text of 10 June 2026, measure 1.1 and sub-measures 1.2.2 and 1.2.3, pages 29-33. Spoken notice, reminders, earcons and Task Force. Primary source.
4. First Draft Code of Practice on Transparency of AI-Generated Content (opens in a new tab), European Commission, 17 December 2025, pages 28-29. The 30-second and five-second rules. Primary source.
5. EU icons for labelling AI-generated content (opens in a new tab), European Commission. Icon next to the player. Primary source.
6. audiowmark (opens in a new tab), Stefan Westerfeld. Free audio watermark. Primary source.
7. Guidelines on the transparency obligations of Article 50 (opens in a new tab), European Commission, C(2026) 5054, page 36: the podcast with cloned voices and the fictional voices example. Primary source.
8. My test on a Drupal 11 test site, with FFmpeg and audiowmark, read with ffprobe. First-hand measurement.