AudioSeal

AudioSeal is Meta's audio watermarking system for proactive detection of AI-generated speech. It jointly trains a generator that embeds an imperceptible watermark signal into a waveform and a detector that returns a per-sample probability of the watermark being present. That localization is the key innovation: instead of a binary watermarked-or-not verdict, it can pinpoint which segments of a longer clip, such as a partially voice-cloned podcast, were synthesized.

Presented at ICML 2024, it uses a perceptual loss inspired by auditory masking and a fast single-pass detector that runs orders of magnitude faster than prior approaches. Meta released the code and weights publicly with a commercial-use license, positioning it as infrastructure for platforms fighting voice-cloning scams and misinformation. Independent evaluations note that, like all post-hoc audio watermarks, it can be degraded by neural codecs and denoisers, so it is a layer of defense rather than a guarantee.