[Discussion]Separation of rhythm guitar and lead guitar

[KilYep]

❓ Questions

First of all, thank you for the great demucs project! I would like to ask if it is feasible to train a new model using demucs that can separate leading guitar and rhythm guitar, and if this could achieve good results. Because I also understand that separating leading and rhythm guitar seems to be an extremely tricky problem. But if it is feasible, how should such a new model be trained? Should it be trained from scratch or based on existing demucs models like 4stems? Approximately how much time would it take to train a model like this?

[CarlGao4]

I've tried to train a guitar sep model, but the results are not good, like though bass can reach a nSDR of 10+, but acoustic guitar can only reach about 3

[KilYep]

I've tried to train a guitar sep model, but the results are not good, like though bass can reach a nSDR of 10+, but acoustic guitar can only reach about 3

Oh, that's too bad. Is the poor performance of guitar sep models due to the fact that separating different guitars is inherently difficult for algorithms, or is it more of a dataset issue?

[johnfalstaff90]

Hello. Can you tell me how did you train the model, maybe I could come up with good dataset.

[CarlGao4]

I used both MedleyDB and moisesDB as the training data I'd like to collaborate with you to try training a new one

[johnfalstaff90]

MedleyDB

moisesDB is something like 80 GB, I don't know how much resources would it take to train, as for MedleyDB, I sent a request to download it but nothing yet and I can't find it anywhere else, I have begun to select from the musdb18 and musdb18-hq songs with "other" stems that contains just lead or rhythm guitar.

[MAMware]

those 80GB how time is? is realyl needed thath much? the issue i foresee when separating distinct roles of the same instruments is not about the sound but its time, i mean its melody. In order to separate both that "melody follow" should be somehow parametrized