Original title: CRAW: Codec Robust Audio Watermarking
David Chernin, Ethan Fetaya · September 2026
Generative speech models make fake voices sound real, creating risks of fraud and misinformation. To fight this, creators can embed an imperceptible audio watermark, which is a hidden signal that proves where speech came from. However, normal file compression and noise cleaners usually wipe these marks out. The researchers created CRAW, a new audio watermarking framework. It works by training the system to expect distortions, using attention-based pooling, hiding the watermark under sounds humans cannot easily hear during playback, and using error-correcting codes to restore lost audio quality. They found that CRAW survives neural codecsSystems that compress and decompress audio using neural networks, often altering the sound., denoisersTools that remove noise from audio, which can also strip out hidden signals like watermarks., and vocodersModels that convert speech representations into audio waveforms, used in synthesis. better than previous methods. At the same time, it keeps the audio sounding just as clear as existing tools.
Imagine writing a secret message in pencil on a piece of paper that will be photocopied several times. Each photocopy makes the text fainter, and eventually, the message becomes unreadable. In this paper, the secret message is like an audio watermark, and the photocopying is like audio compression and denoising. CRAW is a way of writing the message so that it stays legible even after multiple copies, by using special ink and embedding it where it won't be smudged.
As AI-generated voices become indistinguishable from real ones, verifying who said what gets harder. This matters for preventing scams, fake news, and impersonation, where audio proof is crucial.
The abstract indicates that CRAW was tested in experiments, but it doesn't establish performance in all real-world conditions. It maintains perceptual quality comparable to existing methods, not necessarily superior. The focus is on neural re-synthesis transformations, so robustness against other types of attacks isn't addressed.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
CRAW: Codec Robust Audio Watermarking — David Chernin, Ethan Fetaya
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.