Stems for YouTube

Separate any YouTube song into six stems locally with WebGPU. Offline, not live: one 7.8 s-window pass, then instant replays.

Ever loved the drums in a song and wished everything else would get out of the way? This does that, on the original recording. Click once on any YouTube video and Stems separates the song into its individual parts with a neural network running on your own graphics card. Then it plays those parts back in sync with the video, and you decide what is in front. Nothing is re-recorded. You are hearing the original performance, balanced differently. WHY NOT JUST USE AN EQUALISER An equaliser moves frequency bands, not instruments. Two parts can occupy the same frequency range, so turning up "high" does not turn up the piano. This separates the actual parts, which is why pushing the drums forward sounds like more drums rather than more treble. HOW IT WORKS - One pass over the song, about a minute, entirely on your own device - The result is cached, so replays of the same video are instant - Click a part to lift it 9 dB and pull the rest back 6 dB. The whole mix is then levelled to a fixed loudness, so you compare balance rather than volume - Seek, pause, change playback speed or switch tabs - playback stays locked to the video WHAT YOU NEED - Chrome 113 or newer, with WebGPU - Around 7 GB of free memory while it works - A reasonably recent graphics processor. Integrated graphics from several years ago may be too slow. HONEST LIMITS - This is not live monitoring. The model needs 7.8 seconds of audio at a time, so the first pass waits and every replay after that is instant. - Separation is a best guess. Soloing one part can sound hollow where the model dropped a transient. - The "other" track is a bucket, not an instrument: anything the model cannot isolate lands there. - It does not work on Spotify or other DRM services, because their audio cannot be captured. PRIVACY Nothing is uploaded. No account, no server, no analytics, no tracking. The audio never leaves your computer. Open source, MIT licensed: github.com/typosbro/instrument-focus