https://sevag.xyz/free-music-demixer/

free-music-demixer

Open-Unmix (UMX-L) running client-side in the browser with
WebAssembly

View on GitHub

Free AI-based music demixing web app

Upload a song to decompose it into bass, drums, vocals, other, and
karaoke components by applying music source separation (aka music
demixing), powered by a near-state-of-the-art AI model, Open-Unmix,
with the UMX-L pretrained weights. This site is created and
maintained by Sevag H.

[music-demi]

Runs locally in your browser!

Unlike similar products, it's free to use and doesn't store your
data. All processing is done in your browser, and your files are
never uploaded anywhere. It runs well on computers and very slowly on
smartphones; user beware.

Demo

For the demo, I used a free song by Jaxius Music, which you can also
download if you don't have any music files handy.

Here's a demo of 20 seconds of demixed outputs; press the play button
and toggle sources to add or remove them from the mix:

Play now!
[*]Drums
[*]Vocals
[*]Bass
[*]Other

Ready to try it on your own music files?

Demixer app

Download weights (45 MB)
Downloading weights...
[                    ]
Load audio and demix
Demixing progress...
To cancel the running job, refresh the page

Demixing outputs...
Photo by Dylan McLeod

Disclaimers!

  * You can only use the outputs for non-commercial applications as
    per the UMX-L weights license
  * The task is CPU and memory intensive (up to 4 GB), please be
    patient!
  * Long tracks may crash due to the 4 GB RAM limitation; try to keep
    them 5 minutes or shorter
  * Input files can be almost any audio format, but the outputs are
    always stereo wav files @ 44100 Hz

Technical details

The inference code is written in C++, using Eigen3 for numerical
operations. Emscripten is used to compile it to WebAssembly. The
model weights are quantized and compressed from 424 MB down to 45 MB
with a slight hit to performance. View source code on GitHub.

This is a web adaptation of umx.cpp, which is more focused on parity
with the original model. This project was inspired by the "AI at the
edge" GGML project (including whisper.cpp and llama.cpp), and
WebAssembly is a great demo of client-side AI.