Item — HN Reader

MusicLM [1] is the most impressive demo I've seen yet. The problem is we only have access to selected examples (=demo) so we don't know how it performs "on average".

IMO, for music, the limiting factor today is more copyright (=datasets) than models per se.

[1] https://google-research.github.io/seanet/musiclm/examples/

1 comments

The only thing better about these than any other I have heard is that the synthesis quality is better. It just sticks cliches into new contexts and like riffusion and the rest probably has no ability for the longer structures and developing the piece over time in an effective fashion, hence the short samples.

If copyright was the issue than we should have AIs that can effectively do older styles of music which have a great deal of public domain music to work off of, but we don't.

Yeah, the samples are short and "cliche", I agree. Impressive is subjective : creating that from text is already impressive for me.

Also, when I said copyright is limiting, I was thinking about generating "short" and "cliche" emulations of any style of music or artist (the way you do it in early stable diffusion for images)

I might be wrong, but I personally don't expect "innovative developing compositions" anytime soon.