SynthID

ElevenLabs has partnered with Google DeepMind's SynthID technology to embed digital watermarks into audio generated by its AI. The watermark is inaudible to the human ear and persists even after common transformations such as clip trimming, speed changes, metadata removal, and file format conversion.

It is currently being applied first to Text to Speech outputs for free users, and will expand to all ElevenLabs audio within the coming weeks. Watermark detection is available to everyone through the newly launched free webpage ElevenLabs Audio Detector.

In terms of technical requirements, SynthID met all of the following criteria:

• No TTFB (time to first byte) latency
• High detection rate, low false positive rate
• Robustness against transformations such as compression, clipping, and speed changes
• No degradation in audio quality
• Cannot be copied onto audio not generated by ElevenLabs

This move is aimed at strengthening transparency and accountability, and will also be linked with existing C2PA credentials. Going forward, there are plans to add SynthID to the C2PA soft binding list, enabling it to be reattached to audio from which content credentials have been removed.