facebookresearch/flowdec

FlowDec is a tool that compresses audio files to a tiny fraction of their original size while keeping them sounding good. Think of it like a zip file for music or speech, it shrinks a 48 kHz audio recording down to either 7.5 kilobits per second or 4.5 kilobits per second (that's roughly 100-200 times smaller than uncompressed audio), then reconstructs it when you want to listen. The problem this solves is real: audio files are huge, and sending them over the internet or storing them costs money and bandwidth. Existing compression methods either throw away quality or use older techniques. FlowDec combines two modern AI-based approaches, a codec that learns to compress audio intelligently, plus a "stochastic postfilter" that uses flow matching (a newer type of AI model) to clean up any remaining artifacts and make the compressed audio sound closer to the original. You'd use this if you're building a voice chat app, music streaming service, or anything that sends audio over the internet or stores it in large quantities. Instead of sending 10 MB of uncompressed voice, you could send 100 KB instead. The README includes a demo notebook you can run to compress your own audio files using pre-trained models that the authors have already trained and released. The project includes everything needed to both use the pre-trained models and train your own on custom audio datasets. Training requires you to prepare your own data, the authors don't provide datasets, but they give detailed instructions on how to format it and which configuration files to use. The code is built on standard machine learning tools and released by Facebook Research alongside a published academic paper, which is typical for serious research-grade audio projects.