Amazing Project

[BlitzNCS]

Hi

A huge well done on training this model, It works really well and is incredibly useful

I work at a recording studio and I've used drumsep to supercharge my recording workflow - I use it on every drum recording or drum mixing session in some way.

I mainly use it to de-bleed drum stems and it does an awesome job, but it works really well as a reference tool for songs that don't have multitracks available. I use it alongside HTDemucs to really break down reference mixes!

Are you done with this project? I'm really interested in building a similar model or perhaps fine tuning yours, but I'm struggling to set up demucs for training and could use some help! I have access to a massive bank of training data and I'm good with programming concepts, but I don't really understand that demucs training documentation, or maybe I'm missing something.

If you're able to give me a rough guide of how you used demucs to train the model, I would really appreciate that!

Also, since there is now a 6 source Demucs model, as you managed to train this model, do you think it would be possible theoretically to further break down drum sources potentially with more specific training data? I.e kick, snare, toms, hihat, cymbals, percussion (tambourine/shaker)?

Again, thanks for this, it's become a really valuable tool!

[inagoy]

Thank you very very much! I'm glad that you found drumsep useful. I would like to know a bit more about how you use it to de-bleed drum stems!
Setting up and training Demucs can be a bit of a headache, especially if you don't have a powerful GPU. I trained the Hybrid Demucs version, to do it I followed the recommendations in these two Issues: Issue 1 and Issue 2. Also, you might find this doc useful.

In response to the question regarding the separation of more sources, you can set any amount of sources (if you have the corresponding data), but the more you try to separate, the less difference there is between the audio signals, and so the harder it is for the model to discriminate between them. In my experience, hi-hats and cymbals can be very difficult to discriminate against. But I think that a well-trained model with the correct dataset should achieve great results.

If you have more specific questions, you can ask me anything. You can also e-mail me at [email protected].
And yes, I'm currently working on this project, and I want to train a second model with more data and for more epochs!

[BlitzNCS]

Hi, thank you so much for getting back to me! Very much appreciated, I will have a look at those links!

I have sent you an email, just letting you know in case it goes to your junk folder :)

Cheers!

[acosmicflamingo]

Whoa!!! Great work!!! What is the licensing on these models? My perhaps poor understanding is that right now the models that Facebook's hybrid Demucs uses is for educational purposes, which I imagine is why there isn't a single iOS app out there powered by Demucs. However if you created your own models and are now putting it out there for anyone to download, there could be a situation where people are monetizing on your hard work.