I found some discussions online, that SpectraLayers may use AI stem separation models made by the Ultimate Vocal Remover community and allegedly doesn’t has the permission to do so by the original authors.
It would be nice to get a statement from Steinberg regarding this. Is this true or some made-up stuff?
I remember there was an angry comment or post about stolen AI models some weeks ago when SpectraLayers 12 was released. The comment was deleted shortly after.
In the link below there is a discussions about AI separation models like DrumSep-Aufr33-Jarredou or 17_HP-Wind_Inst-UVR, which apparently can be found in clear text inside the SpectraLayers executable.
Ok, sadly it seems to be true that SpectraLayers incorporates models from other developers without any attribution whatsoever.
Robin writes in the topic that they use such models, I found this comment on page six.
It would be good for transparency to attribute the original developers because they trained the original base models, sometimes even with their own datasets.
As I understand it, Steinberg might have further “refined” some of the models, but I don’t think that frees them from an ethical way to attribute the original authors and name the original models.
Yeah, but the problematic drum separation model wasn’t released under any license. Then it’s not like you can simply take it, which SpectraLayers did.
This was the problem the original author had. They didn’t ask for permission. If they had asked, they very likely would have gotten the okay and maybe even a better trained version, like mentioned in the thread above.
The current license is CC-BY-NC-SA as you can see here:
Firstly, I am always amused that anyone would ever think that anything made freely available on the internet could ever be protected in any way – why do you think Microsoft bought GitHub?
Secondly, while I absolutely support the concept of correctly attributing credit where credit is due to the efforts of developers everywhere, especially open source contributors, I don’t think it’s fair for @Sophus to come on here, make what is an accusation, and then mark their own response as the “solution”. In the Audiofanzine topic referred to (and French is not my native language) Sophus has omitted to mention a response from @Robin_Lobel himself where he attempted to contact the developer and did not get a response.
Thirdly:
… so, it was put on the internet with no license information specified?
If it was put on the internet with no licensing conditions specified, I’m afraid yes, anyone can and will take it – that might not be moral, but not necessarily illegal.
And finally, @Sophus , please explain what your motivation was to start this thread, for example, how does this affect you (or any of us) – is it merely a moral discussion, or are you a lawyer, a researcher of AI models or open source, or party to some legal proceedings, or maybe you know (or are) the guy who made the “angry comment or post about stolen AI models some weeks ago when SpectraLayers 12 was released” ?
I’m sure that account being created the very day SL12 was released, who just happens to remember a deleted post, and just happens to solve their own post of accusations without proof or even waiting more than 6 hours for anyone to actually discuss it is just a coincidence, man! All just coincidence…
He tried to contact the developer AFTER the first accusations appeared, on page three of the discussion. Link is here:
The solution is a link to an response from Robin, in which he basically answered my question, that SpectraLayers incorporates other AI models available on the Internet. I hadn’t translated everything there before I posted my question.
Besides that, I think it’s important to know for others. I didn’t even know that SpectraLayers uses UVR models. I thought they used Spleeter in the beginning, then switched to Demucs and finetuned that over the years.
Now I know they use modified UVR models.
I’m pretty sure this is not how European copyright law works.
If SpectraLayers uses models from the Internet it would be nice to know which one, so I can:
download them myself
get updates earlier from the original authors instead of waiting another year until the next version of SpectraLayers
use them in UVR with working GPU acceleration and don’t need to wait hours for processing in SpectraLayers while getting basically the same results
I think this was my original response to one of the rant comments some weeks ago, where the commenter stated something about models from UVR. The comment I responded to is no longer there.
My plan is basically to find a way to get the best possible stem separation.
In UVR you can combine different models in an ensemble mode, where it runs the audio through each model individually and after that automatically combines the individual results via a min-max-average summing. This can lead to much improved results.
I tried to do the same in SpectraLayers but you cannot do batch unmixing with different settings because the resulting layers always have the same name and then the batch process stops. Doing it manually for each setting is very time consuming, especially because the current processing in SL 12 is slow.
One question for me is, if it makes sense to invest time here to experiment, if the models are nearly identical to UVR.
Why not just use UVR then, or code your own solution?
I am just trying to understand what problem it is that you were trying to solve by creating this topic, which seems to be suggesting something malicious or inappropriate had been done.
For me it’s to expensive. You need to rent GPUs for that and that is going to cost triple to quadruple digits to train or finetune one model.
I don’t know all the original models. I think the drum separation is based on the DrumSep-Aufr33-Jarredou model and the sax model might be based on the 17_HP-Wind_Inst-UVR model.
What are the remaining models based on? If SpectraLayers would attribute them, I should be able to check them out.
If I had a chunk of the code for Cubase, I could build my own daw.
Geez dude….go do your own research on your own dime. If you can’t afford systems and don’t have the ability to sift through every algorithm on the planet to find what works best on your sources, that’s your problem.
Yes, I know. My comment was regarding your “code your own solution”. I think you know what I was talking about.
I already own the newest SpectraLayers. I’m using it since version 4. And I compared the models inside SpectraLayers with UVR. At least the Bs-Roformer models in UVR create much better instrumental stems than what SpectraLayers can do. Highly recommended, if you also want to try it out on your own.
What’s happening under-the-hood is more complex and nuanced than what the original plaintif seemed to suggest, I tried to give hints at that in my answer (such models are not just re-used as-is), but given his very aggressive attitude from the start it was not possible to have a nuanced discussion about it, specially in a public forum.
Even regarding his initial complain there’s more nuance than what he implied. But he never replied to my PM, and seemed to me he was not open to a civil discussion about it anyway. To settle this significant changes to the Unmix Drums module have been made, which is a win-win for all parties : separation quality was improved overall from the initial SL release, more drums stems can be unmixed from the initial SL release, users win from this, and plaintif was satisfied with the change.
I’ll close this topic as it seems enough has been said on that matter…
