Voluntary model and data/training set contributions from users?

SpectraLayers
1

There’s obviously a big contention with companies in the background gleaning data and models off peoples work without them knowing/missing some fine print somewhere,

But I’m sure some people have some material, samples, or even the time to record new models to give voluntarily for the sake of improving SL models? Perhaps also with the caveat that anything provided wouldn’t be used for music generation models by Steinberg/Yamaha.

The bigger the training set, the better? no?

2

I’m sure others would do this but it has to be for a very good price. The problem with a lot of companies is there is way too much greed and exploitation.

Back in Albert Einstein’s days, he would publish his theories in papers and many physicist would buy them, however he would also do a lot of free lectures and was known to mentor a lot of physicist students (out of a good heart). Now (2026) someone on Albert Einstein’s is not able to do anything for free because the economy won’t allow it. There’s people out there who are eager to share knowledge (at no cost) just for the sake of improving the world, but (as of right now) that is not possible!

3

SL already takes open source models from the Internet. There was even a fallout with the creator of the drum splitting model because they didn’t ask for permission to use the model in SpectraLayers. Was mostly talked about on French sites, so not very public. This problem was later resolved.

You can find a lot of these models here and also a training software to improve them, but it’s expensive because you usually need to rent a server for fine-tuning. This can cost multiple hundreds and up to thousands of dollars to fine-tune a model.

Music-Source-Separation-Training/docs/pretrained_models.md at main · ZFTurbo/Music-Source-Separation-Training · GitHub Here are the models

GitHub - ZFTurbo/Music-Source-Separation-Training: Repository for training models for music source separation. Here is the software

MUSDB18 | SigSep Here is a dataset for training

MVSEP - Music & Voice Separation Here you can find some datasets for testing