EU AI Act: watermarking for AI-generated audio

The EU AI Act requires AI-generated audio to be marked in a machine-readable way and detectable as artificially generated. Our patented audio watermarking embeds a robust, inaudible mark into synthetic speech, voice clones, and music at the moment of creation — and lets you, your partners, and platforms detect it reliably later.

Request information or evaluation

What the EU AI Act requires

  • Providers of generative AI (Article 50(2)): providers of AI systems, including general-purpose AI systems, that generate synthetic audio, image, video, or text must ensure that the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. The technical solutions must be effective, interoperable, robust, and reliable, as far as technically feasible.
  • Deployers (Article 50(4)): those who use AI to generate or manipulate audio constituting a "deep fake" must disclose that the content has been artificially generated or manipulated.
  • Timing: these transparency obligations apply from 2 August 2026.
  • Scope: the rules apply to AI systems placed on the EU market or whose output is used in the EU — regardless of where the provider is based.
  • Techniques: the Act's recitals explicitly name watermarks, metadata, cryptographic methods, logging methods, and fingerprints as possible marking and detection techniques.
  • Penalties: non-compliance with these obligations can lead to fines of up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher.

Why audio watermarking — and not only metadata

Metadata and file tags are lost as soon as audio is re-encoded, streamed, edited, re-recorded, or played through a loudspeaker. A watermark is part of the sound itself, so the mark travels with the audio wherever it goes:

  • Robust — survives lossy compression (MP3, AAC, streaming platforms), cutting and editing, and, with AWT5, broadcast and loudspeaker-to-microphone transmission.
  • Inaudible — the quality of your generated speech or music is not affected.
  • Machine-detectable — detection is automatic and fast, suitable for large-scale screening and moderation.
  • Informative — the watermark carries data, not just a yes/no flag: for example, the provider, model, and generation time.

How it works

  1. Embed at generation. Integrate the AWT SDK into your text-to-speech, voice-cloning, or music-generation pipeline, or run the command-line tools on your servers. Real-time, low-latency stream encoding is supported on all major platforms, including WebAssembly.
  2. Choose the payload. Encode the information you need — for example, a provider ID, model version, generation timestamp, or content/session ID (up to 120 bytes with AWT2, up to 20 bytes with AWT5*).
  3. Detect anywhere. Decoder tools and the SDK let you, your partners, or platforms verify whether audio was generated by your system and read the embedded data — or build your own public verification service.
  4. Go further (optional). Mark human-created recordings as authentic, and register fingerprints of generated content with AFT as a complementary identification layer.

(*) subject to limitations in different packages

Which technology fits

AWT2

Files and digital streams. Blind, highly robust watermarking with a large payload — ideal for marking generated audio delivered as files or streamed through digital channels. More about AWT2 →

AWT5

Broadcast and over-the-air. Designed to survive loudspeaker playback, broadcast paths, and room reverberation — for generated audio that may be aired, played in public, or re-recorded on a phone. More about AWT5 →

AFT

Complementary fingerprinting. Register acoustic fingerprints of generated content to identify it by its sound, as an additional layer next to watermarking. More about AFT →

Who it is for

  • Text-to-speech and voice-cloning providers
  • AI music and sound generation services
  • Voice assistants and conversational AI
  • Platforms and content moderation teams
  • Broadcasters and publishers using synthetic voices
  • Enterprises deploying generative AI for audio

Why work with us

  • Patented, proven technology — developed since 2007 and used by broadcasters, labels, publishers, and technology companies (see customers, patents).
  • Runs on your infrastructure — the software is licensed and deployed on your own servers or devices; your generated audio never has to leave your systems.
  • Integration-ready — command-line tools for automation, and a thread-safe real-time SDK for Windows, Linux, macOS, ARM, iOS, Android, and WebAssembly.
  • Individually prepared evaluation packages and direct technical support from the inventor and developer of the technology.

Preparing for the EU AI Act? Tell us about your generation pipeline and target platforms. We will answer your questions and discuss the next steps toward an evaluation.

Evaluation packages are prepared individually, following an initial discussion of your use case and requirements.

Request information or evaluation Read the article on AI and audio authenticity

This page provides general information and is not legal advice. Please consult your legal advisers regarding your specific obligations under the EU AI Act (Regulation (EU) 2024/1689).