Meta has published research on a system for embedding data in sound. This signal is imperceptible to humans but can be detected by a system. It works similarly to watermarking in images to extract data. Meta plans to integrate this system into AI-generated audio files.
Known as AudioSeal, this tool differs from traditional data embedding systems like watermarks, which only insert data at specific intervals, making it easy to remove. AudioSeal inserts data at every point of the audio file, even in files that are hours long.
The current challenge in the industry is the lack of standards for watermarking in audio files, especially to determine if a file was AI-generated. Decoding and verifying such files, especially long or edited ones, can be time-consuming. However, AudioSeal’s method is currently achieving an accuracy level of 90%.
Meta has disclosed the workings of AudioSeal on GitHub.
Source: MIT Technology Review, Image from Pixabay
**TLDR:** Meta introduces AudioSeal, a system for embedding data in sound, allowing for imperceptible data detection. This tool differs from traditional methods by inserting data at every point in an audio file, making it harder to remove. The challenge lies in establishing standards for verifying AI-generated audio files. AudioSeal achieves a 90% accuracy rate in its current implementation.
Leave a Comment