Detecting audio generated by ElevenLabs with SynthID

People should know when they are interacting with AI. In the past, AI-generated content was easy to spot – it sounded robotic or had six fingers. But with recent improvements in models, it is getting much harder to tell.

As our voice, music, and sound effects models improve, we want people to be able to identify if audio was generated by AI without specialist tools. That’s why we’re partnering with Google DeepMind on SynthID: a digital watermark embedded directly into audio generated by ElevenLabs. These watermarks are inaudible to people, and remain even when clips are trimmed, sped up, stripped of metadata, or converted into a different file type. This week, we have started including SynthID in Text to Speech generations by free users, and we will expand coverage to all ElevenLabs audio generations over the coming weeks. Importantly, these watermarks will be detectable using our new free ElevenLabs Audio Detector. 

Bolstering transparency and accountability

We have long prohibited the use of our tools to deceive, manipulate or harass people, and our systems are already designed to trace content back to the user who generated it, so we can take appropriate action. Watermarking with SynthID advances our commitment to transparency and accountability by allowing members of the public to verify the source of an audio clip.

In some cases, people will simply want to know whether or not content is AI-generated. In others, it will be important to understand which AI platform it came from. That’s why we are today launching our own free ElevenLabs Audio Detector webpage that allows people to verify whether audio was generated by ElevenLabs. This builds on our existing AI Speech Classifier, but uses SynthID to more robustly embed attribution directly into the audio. This mechanism is important to keep us publicly accountable for the power of our technology and take appropriate action if a malicious actor manages to defeat our comprehensive safeguards and create convincing deepfakes. 

A growing number of jurisdictions require AI-generated content to be marked as synthetic in a machine-readable format. Watermarking with SynthID complements our existing ecosystem of provenance and compliance tools, which also includes C2PA credentials. We are also excited about the possibility of adding SynthID to the C2PA soft bindings list so that audio stripped of its content credentials can have them reattached. 

High-quality watermarking solution

SynthID works by hiding a pattern of sound in audio clips. The pattern is inaudible to the human ear but detectable by our ElevenLabs Audio Detector. Each audio file gets its own unique pattern, and these patterns survive common audio transformations like compression, clipping, and speed changes.

SynthID performed well in our benchmarks and hit all of our technical requirements:

• No added time to first byte (TTFB) latency 
• High detection rate with a low false positive rate
• Robust to cropping and other transformations that are commonly encountered online
• Imperceptible to the human ear, with no audible quality degradation
• Cannot be copied onto audio that ElevenLabs did not generate

We look forward to continued collaboration with the SynthID team at Google DeepMind to advance state-of-the-art audio watermarking.

What’s next

Watermarking is primarily about transparency and accountability, but it also enables new product capabilities.There is already an established market for digital content credentials – studios use them to protect intellectual property, and individual creators use them to ensure they are paid when their work is reused. In the future, watermarks might enable creators and IP holders to embed their own metadata directly in content, allowing them to detect and act on copyrighted material redistributed on platforms like YouTube, Instagram, and TikTok.

Watermarking is one part of our broader transparency commitment. As our models get more powerful and lifelike, our accountability infrastructure needs to keep pace. The more that we can all attribute content to its source, the more trustworthy our shared information ecosystem will be.