Beginner's guide to getting free stems with free-music-demixer
Posted by sevagh on 23 September 2023
Category: getting-started
Table of contents
- Prepare your music files
- Navigate to this site
- Download weights
- MAX MEMORY option and multi-threading
- Demixing a single track
- Demixing a batch of tracks
- Canceling a running job
- Reporting bugs
Hello! I’m the creator of this site, and this post describes how I use my own website to get stems from mixed songs for free.
Last updated: 2023-12-29
Prepare your music files
Let’s cover two scenarios:
- Single track: a single audio file containing the mixed song you want to demix
- Batch of tracks: a folder containing multiple songs you want to demix
Most common audio file extensions (wav, mp3, opus, flac, webm) should work. Note that the output stems are always returned as stereo wav files with a 44100 Hz sampling rate.
Navigate to this site
When you visit https://freemusicdemixer.com, you land on the homepage of this site. From this blog post, you can also click Home in the top bar. Once you’re on the home page, scroll down to the section named “Demixer apps”, or go there directly:
Download weights
Decide which model you want first.
- “Demucs 4-source” is the default model “htdemucs” of the Facebook Demucs v4 hybrid transformer model
- Outputs 4 stems: vocals, drums, bass, melody (+ instrumental)
- “Demucs 6-source” is “htdemucs_6s”,
- Outputs 6 stems: vocals, drums, bass, piano, guitar, other (+ instrumental)
- “Other” generally contains the melodic track remaining after removing piano and guitar
Enable the “Show dev logs” checkbox to display developer messages on the screen to describe how the model weights were loaded:
Both apps are now ready to use.
MAX MEMORY option and multi-threading
When Demucs runs on a track, it consumes a max of 4 GB of memory. Demucs can be applied to independent segments of the track, which is an easily parallelizeable workload. We can split a song into N segments, and process it with N workers (that consume a max of 4*N GB memory).
You can now choose up to 32 GB of memory consumption to launch 8 workers. This can roughly be considered as launching 8 threads or processes in your browser, which should result in dramatic speedups (on my tests, it took a 17 minute track down to 5 minutes).
Note: from my tests, up to 32 GB works in Firefox, but only 16 GB in Chrome. I need to run more tests on Chrome to track down the issue (for now I reported this issue, possibly in the wrong place).
Demixing a single track
In the “Single track” app (the first one), click “Choose file”, browse to your song file, and then click “Load audio and demix.” As per the above, you can raise the memory limit to speed up the processing.
As the demixing proceeds, the progress bar will start filling up:
In the dev logs, on the left pane for Javascript messages, you will see the demix job for your track being kicked off:
[Javascript 15:44:33] Selected model: demucs-4s
[Javascript 15:48:16] Initializing 4 workers!
[Javascript 15:48:16] Beginning demix job
[Javascript 15:48:18] Worker 0 is ready!
[Javascript 15:48:18] Worker 1 is ready!
[Javascript 15:48:18] Worker 3 is ready!
[Javascript 15:48:18] Worker 2 is ready!
In the right pane for C++ messages, you will see the actual steps of the AI model being executed:
[WASM/C++ 15:48:41] (WORKER 3) Time (crosstransformer): layer 4
[WASM/C++ 15:48:41] (WORKER 3) Crosstransformer: finished
[WASM/C++ 15:48:41] (WORKER 1) Time (crosstransformer): layer 4
[WASM/C++ 15:48:41] (WORKER 1) Crosstransformer: finished
[WASM/C++ 15:48:41] (WORKER 3) Freq: channels downsampled
[WASM/C++ 15:48:41] (WORKER 3) Time: channels downsampled
[WASM/C++ 15:48:41] (WORKER 1) Freq: channels downsampled
[WASM/C++ 15:48:41] (WORKER 1) Time: channels downsampled
[WASM/C++ 15:48:42] (WORKER 0) Freq: decoder 0
[WASM/C++ 15:48:42] (WORKER 2) Freq: decoder 0
When it finishes, the final messages printed in the C++ log pane are:
[WASM/C++ 15:48:55] (WORKER 3) padding offset is: 72765
[WASM/C++ 15:48:55] (WORKER 3) Copying waveforms
Finally, on the Javascript log pane, there will be a finished message:
[Javascript 15:48:55] Summing segments
[Javascript 15:48:55] Normalizing output
[Javascript 15:48:55] Preparing stems for download
In the app itself, at the bottom under “Demixed outputs,” you will find the stem files available for download:
There are the demixed stems from your track! Enjoy.
Demixing a batch of tracks
In the “Batch demix” app (the second one), click “Choose file”, browse to your folder of tracks, and then click “Start batch demix,” where you will be asked if you accept to upload all of the files:
N.B.! I use the word “upload” but there is no uploading being done anywhere outside of your computer. A more accurate term is “load”, since it’s your web browser that is locally loading your audio file.
After you accept, the progress bar will fill up as each entire song is finished. The Javascript dev logs for the batch job will show the progress:
[Javascript 16:04:07] Initializing 8 workers!
[Javascript 16:04:08] Beginning batch job for shimmy_shimmy_ya.wav
[Javascript 16:04:13] Worker 0 is ready!
[Javascript 16:04:14] Worker 1 is ready!
[Javascript 16:04:14] Worker 3 is ready!
[Javascript 16:04:14] Worker 2 is ready!
[Javascript 16:04:14] Worker 4 is ready!
[Javascript 16:04:14] Worker 5 is ready!
[Javascript 16:04:14] Worker 7 is ready!
[Javascript 16:04:14] Worker 6 is ready!
When the entire folder is done, you will have links to download zip files for each song in the folder containing the demixed stems per song:
Canceling a running job
At any point if you want to cancel the current run, just reload the website. You will have to click the weights download button again, but this typically won’t require a redownload because the weights file will be stored in your browser cache.
Reporting bugs
If any of the above steps don’t work or result in strange outputs, I invite you to open a bug report on the project’s GitHub Issues or e-mail me directly (contact at freemusicdemixer dot com).
