Hacker News Reader

Hacker News

Comment by justinclift | Hacker News Reader

Parent Full thread

justinclift

Wonder how far off the whole "generate music based on your existing music library" thing is going to be?

That'll make musicians happy with big tech as well, just like artists are. *sigh*

View on HN

The Record labels are far, far more litigious than the art community.

They can't litigate a person doing this at home, and never redistributing.

I suppose they might try, anyway.

The RIAA pioneered copyright enforcement at the individual level back in the 2000s, they absolutely would try to sue downstream AudioCraft users.

They should start with AudioCraft itself, conceptually it's derivative work and it doesn't matter if it's "open source" or not. Try throwing in someone's sample in a song and publish it saying "no copyright infringement intended and I totally don't make any money from it"... If it becomes popular, see how long it stays up until DMCA takedown. And we know this dataset is already popular.

> and publish it

This is precisely the opposite of the context I was remarking on.

AudioCraft itself is published. That's the context I am remarking on.

Is training a "pirate model" something you'd reasonably be able to do at home though, given the compute requirements? The analogous "image generation at home" is only possible due to a for-profit entity with significant resources choosing to (a) play fast-and-loose with the provenance of their training set and (b) giving away the resulting model for free, if the open source community had to train their models from scratch then as best as I can tell they would still be stuck in the dark ages generating vague goopy abominations.

Currently, yes, available compute power @ home does indeed seem like a limitation. Whether that remains true going forward seems a little unclear to me.

You could take a model trained on CC content and then fine tune it on copyrighted material cheaply and quickly

Ugh, I dread having to listen to everyone's hyper personal music because they swear up and down to the point of tears that "IT'S THE BEST SONG EVER CREATED! EVER!!!", while the constantly prod for you to affirm how amazing the song is.

Bruh, music is subjective as hell, and I can already tell I hate this song.

Perhaps LoRA (Low-Rank Adaptation) training techniques could be used for these types of models, like they're currently being used with LLMs and latent text-to-image diffusion models.

Sadly looks unlikely if the base model wasn't trained on vocals.

> Mitigations: Vocals have been removed from the data source using corresponding tags, and then using a state-of-the-art music source separation method, namely using the open source Hybrid Transformer for Music Source Separation (HT-Demucs).

> Limitations: The model is not able to generate realistic vocals.

(https://github.com/facebookresearch/audiocraft/blob/main/mod...)

I suspect this was a combination of playing it safe and that the model isn't well architected to reproduce meaningful vocals.

Why not do generate music you like which wouldn’t need you to upload your library and would have RLHF baked in.

Something like the algorithm TikTok uses. First probing by offering a variety of content that should match based on what little information you have on the user (ip location, locale, etc).

Then use the user’s action to iteratively refine your classification, until you end up with something tailor-made.

Uh more like Reddit with up and down and also how long you listen.

Anyone having listened to this MusicGen's output samples would surely answer "a million miles".

Seriously, you couldn't sell this output for a free mobile clicker game.

Reply on news.ycombinator.com