Demucs 3
for 4 stems
(SDR 7.7 for 4 stems, it’s better than Spleeter (which is SDR 6.5-7), or better than MDX 4 stem. In most cases, it’s even better than Audioshake - at least on tracks without leading guitar)
Accompanied by MDX-UVR 9.7 vocal model, it gives very good 4 stem separation results
(For Demucs 4 a.k.a "htdemucs" check below)
https://colab.research.google.com/drive/1yyEe0m8t5b3i9FQkCl_iy6c9maF2brGx?usp=sharing (by txmutt), alternatively with float32 here
Or https://huggingface.co/spaces/akhaliq/demucs
Pick up from the list Demucs Model B there.
You can export result files in MP3 320kbps, WAV and FLAC. File limit is 100MB and has a 10 minute audio length limit.
To use Demucs 3 locally: https://discord.com/channels/708579735583588363/777727772008251433/909145349426384917
Currently, all the code uses now main branch which is Demucs 4 (previously HT) but these Colabs use old mdx_extra model.
Demucs 3 UVR models 2 stem only available on MVSEP.com or in UVR5 GUI (nice results in cases when you suffer vocal bleeding i regular UVR5, GSEP, MDX 9.7 - model 1 less aggressive, model 2 more destructive, model bag has more bleeding of all three).
In Colab, judging by quality of drums track, I prefer using overlap 0.1 (only for instrumentals), but default set by the author is 0.25 and is better for sound of instrumental as a whole.
But it still provides decent results with instrumentals.
Also, HV had overall better separation quality results using shifts=10, but it increases separation time (it's also reflected by MVSEP's SDR calculations). Later we found out it can be further increased to 20.
Also, I have a report that you may get better results in Demucs using previously separated instrumental from e.g. UVR.
Anjok’s tip for better instrumentals: “I recommend removing the drums with the Demucs, then removing the vocals and then mixing the drums back in”. Yields much better results than simple ensemble.
It works the best in cases when drums get muffled after isolation, e.g. in hip-hop. You need to ensure that tracks are aligned correctly. E.g. if you isolate drumless UVR track, isolate also regular track to align drumless UVR track easier with drums track from Demucs, otherwise there will be hard to find the same peaks. Then simply align drumless UVR the same as regular track is aligned and mute/delete UVR regular (instrumental) track.
Be aware! This is not a universal solution for the best isolation in every case. E.g. in tracks with busy mix like Eminem - Almost Famous, the guitar in the background can get impaired, and so even drums (UVR tends to impair guitars in general, but on drumless track it was even more prevalent - in that case normal UVR separation did better job).
Also, if you slow down the input file, it may allow you to separate more elements in the “other” stem.
It works either when you need an improvement in such instruments like snaps, human claps, etc.
Normally, the instrumental sounds choppy when you revert it to normal speed. The trick is - "do it in Audacity by changing sample rate of a track, and track only (track menu > rate), it won't resample, so there won't be any loss of quality, just remember to calculate your numbers
44100 > 33075 > 58800
48000 > 36000 > 64000
(both would result in x 0.75 speed)
etc.".
Also, there's dithering enabled in Audacity by default. Might be worth disabling it in some cases. Maybe not, but still, worth trying out. There should be less noise.
BTW. If you have some remains of drums in acapella using UVR or MDX, simply use Demucs, and invert drums track.
“The output will be a wave file encoded as int16. You can save as float32 wav files with --float32, or 24 bits integer wav with --int24” it doesn’t seem to work in Colab.