Is it ethical to train AI voicebanks with copyrighted songs if I want to use them commercially?

[elizaditton]

Hi all, long time no see. (And sorry if I'm posting in the wrong section, I wasn't exactly sure where this question would fit...)

I've been wanting to create a DiffSinger voicebank for a while now, and most if not all the people I've seen make one have covered other people's music for their wav data. For me, this brings up an ethical dilemma. Is it okay to use music that belongs to someone else (even if I'm the one singing) to train a voicebank? Especially if I plan to use said voicebank for commercial purposes? From what I understand (although I could be wrong since I'm still learning), the actual wav data as recorded doesn't end up in the final voicebank, but despite that, if the wav data we're using for training a voicebank uses copyrighted songs that don't belong to us, would doing anything commercial with that voicebank be considered copyright infringement? Since the copyrighted material is kind of being used as a base to go off of by the AI and thus the voicebank?

I don't know, am I overthinking this? What do you all think?

 

[Kiyoteru]

You might want to try looking at it this way. Would the copyright holder have the grounds to sue you for your final commercial project?

 

[LotteV]

Training off of copyrighted songs is perfectly fine, even for commercially released AI banks. They've done that with Synth V as well from what I was told (by a dev no less), so it should be fine as well for things like NNSVS/ENUNU and Diffsinger.

The thing with AI voicebanks is that you can't disassemble them to check the recorded data, since there are no actual audio files in the model itself. As a result, it's really hard to prove a copyrighted song was used for training (so, I've essentially answered the above question for you, lol). This isn't like AI art where an art piece that the AI was trained on can still be recognizable in the end product (which is a problem that does exist when you make an AI bank off of a celebrity's voice, on the other hand. But that's kind of a given). Heck, I've even seen public datasets featuring copyrighted songs and they're yet to be struck with DMCA or the like (of course, if you're scared of this happening anyway, you can always just keep your dataset private just in case).

For a long time, I was also wondering about this exact dilemma, since I plan on creating a Diffsinger bank one day. But seeing as this is basically standard industry practice, I don't see how it could be a problem.