Google and ElevenLabs Adopt SynthID for AI Audio Watermarking Google and leading AI audio firms are racing to make synthetic sound more traceable, even as the underlying technology becomes harder to distinguish by ear. Their answer so far: invisible watermarks baked directly into the audio itself.

The shift began with Google DeepMind’s decision to embed its SynthID watermarking system into audio created by its Lyria music model and YouTube’s new AI audio-generation tools. The company says the watermark is “inaudible” and “doesn’t compromise the listening experience,” while remaining detectable even after compression, speed changes, or added noise. SynthID works by converting an audio wave into a two‑dimensional visualization of how its frequency spectrum evolves over time—an approach DeepMind describes as “unlike anything that exists today.”

This audio rollout followed an earlier beta of SynthID for images generated by Google’s Imagen model on Vertex AI, where similar invisible marks are designed to survive cropping and resizing, though DeepMind concedes they are not foolproof against “extreme image manipulations.” The broader push aligns with policy pressure: watermarking tools are increasingly framed as safeguards against generative AI misuse, and are explicitly encouraged in U.S. policy such as President Joe Biden’s AI executive order.

Soon after Google’s audio announcement, AI voice platform ElevenLabs moved to adopt the same standard, declaring: “ElevenLabs rolls out SynthID support.” The company has added SynthID to text‑to‑speech generations for its free users first, with plans to extend coverage “over the coming weeks” to all audio produced on its platform, and to make those marks detectable via an in‑house ElevenLabs Audio Detector.

Together, Google and ElevenLabs are testing whether a shared, largely invisible layer of provenance can keep pace with rapidly improving synthetic audio—a tool many technologists see as necessary, but not yet sufficient, to curb deepfake harms.

Continue reading https://foxvector.com/stories/019f04f6-e30b-34ef-73ee-3e0e65121221