Do you use specific sample for separation test? I mean, is there like something standard song/track that is difficult to separate? Also any tips on drums only separation is welcome.

 

1 song is not enough, MVSEP people have made 2 datasets to evaluate the different algos : https://mvsep.com/quality_checker/

 

A guy i Know ЌǻÖ is doing Instrumentals with the Group I2U¥
they use UVR5.5

https://www.youtube.com/@i2u26/featured

 

oh give UVR5 a try you can mix and match algos. it also runs locally so you can do it offline or batch convert.

 

since were on this stem separation topic. has anyone found anything that can pull specific sounds out of stems? Like kicks or snares? is the AI there yet? I know it will be soon

 

Do UVR5 for drums with Htdemucs_ft alone then import into RipX. But youll end up with full tracks when you can actually cherry pick any ''best kicks'' or Best Snare amongst the resulting fro UVR5 htdemucs_ft.

Id recommend cherry picking on the Htdemucs_ft result as there is no need to get a full 32 bit float kick line of 4 minutes long. Get thos kick and counter kick and you should be gold.

 

It should be possible to train a MDX models with a custom "drums" dataset. With the full drums mix as mixture, and snare, kick, toms (hihat ?) as individual stems.

I've already seen a MDX model trained to remove reverb on vocals and it was not bad !

 

I do the Htdemucs_ft on drums and its amazing! I import it into Ableton, I only extract the necessary sections. However, sometimes there may still be some cymbals or percussive sounds that overlap. Although this is not a major issue, I am still interested in exploring other available options such as RipX.

Regarding the MDX custom dataset, I have been thinking about the possibility of using a similar approach to that of loopcloud and other sample organization software, which can identify individual instruments quite well. It would be interesting to see if this logic could be applied to an MDX dataset, allowing for the extraction of specific sounds, similar to how loopcloud can tag sounds.

 

There's a thread about drums separation on Demucs' github page with an implementation of NMFToolbox :
https://github.com/facebookresearch/demucs/issues/422

Here is a more recent and better version of the NMF process (but for MatLab)
https://github.com/aaron985/dual_channel-NMF-for-drum-separation/

 

Which is why I specifically detailed the need to cherry pick your kick and snares. You'll find at least a dozen of those per tracks. Then you'll need to EQ them anyways to make them fits your specific production condition. That should be enough to get you in trouble in the sampling rights territories as it will be very recognizable.

 

Ok. Someone already made a Demucs model to separate drums elements ! :D
https://github.com/inagoy/drumsep
The Google colab is working if you download the model manually and put it in the model folder (the function from the code is broken)

No info about the dataset used to train the model...

EDIT : Dataset infos from the dev :
It's trained with 7 hours of drum tracks that I made using sample-based drum software like Adictive Drums, trying to get as many different-sounding drums as I could. As everything was controlled with MIDI, I could export the isolated bodies: kick, snare, toms (all on one track), and cymbals (including hi-hat). So every dataset example is composed of kick, snare, toms, cymbals, and the mixture (the sum of all of them).

 

Last edited by a moderator: Mar 2, 2023

There's also the "Zero Shot" algo that is maybe also able to separate drums elements. It works with 2 audio files, one is the song you wanna extract something, the second is the reference stems you want to extract (for example a short bass recording if you wanna extract the bass from the song). It can work with any source so it should be possible to feed it with drum elements to extract them from a full drum mix. There's an instance running on replicate : https://replicate.com/retrocirce/zero_shot_audio_source_separation

 

Last edited by a moderator: Mar 2, 2023

https://audiosex.pro/threads/polyma...into-a-music-production-sample-library.69171/

 

Interesting - had a quick play around with this last night. I'd be using RIPx for everything. Ripx does a quicker more balanced result but if you want a cleaner vocal then wacking an ensemble on and setting it to all the vocal algos definitely seems to yield better results (though it will be processing for a much much longer time)

 

I think UVR's ensemble mode is only merging wavs output from different models, if I'm right, you can use different algos on mvsep.com and then merge the wav files in your daw (and even balance them as you want). The only downside is that you can't tweak the model parameters (shifts, overlap...)

 

Last edited by a moderator: Mar 2, 2023

let me know if you figure this out.. It currently merges realtime while it is rendering. so the only way I could think about doing it is have the files from mvsep renamed and ready to replace the files generated by UVR before the process is over.

 

So... it doesn't merge the audio while rendering, it renders each model's output one by one, and once all models are done, it merges them with a chosen algos (max/min, min/max, avg/avg, this last "average" one seems to always give better and more stable results).

And I've just realised that UVR5.5.0 has in fact a builtin "manual ensemble" feature in the "audiotools" process method, so you can extract stems with mvsep.com (or any other site) and then "ensemble" them locally with UVR. It's fast and no CPU/GPU explosion ! ;)

 

Last edited by a moderator: Mar 8, 2023

Mind blowing side-note : you can also train and use Demucs to REMOVE DISTORTION from recordings, research paper and sound examples here (the code and models are not public (yet?) unfortunately) https://joimort.github.io/distortionremoval/

 

holy shit!! thank you!! A manual ensemble is what we needed!!