Meta has developed a way to embed watermarks into speech created by artificial intelligence.

Meta has developed a new system called AudioSeal that allows hidden signals, known as watermarks, to be embedded into audio clips created by artificial intelligence.

Here's What We Know

The new technology could help combat the spread of misinformation and fraud using voice cloning tools. For example, it can indicate which parts of the audio in an hour-long podcast were generated by AI. Although watermarking is not yet widespread and there is no common industry standard, the Meta team has achieved watermark detection accuracy between 90 per cent and 100 per cent, a significant improvement over previous attempts. At the same time, Meta does not yet plan to apply watermarks to audio created with its tools. In addition, watermarks for AI-generated content can often be easily tampered with or removed.

AudioSeal is available on GitHub and can be used to add watermarks to AI-generated audio clips. In the future, it can be integrated into AI-audio generation models to automatically apply it to all generated speech. Work on this system will be presented at the International Machine Learning Conference in Vienna, Austria in July.

Source: TechnologyReview