Hacker News new | ask | show | jobs
by CrypticShift 1284 days ago
> written, composed, and (deep-faked) performed

Ok, so :

1- Written : ChatGPT already spits out pop-hit-ready lyrics (not a high bar...)

2- Composed : AI results are, by definition, perfectly attuned to our tastes. Plus pop melodies are the easiest. So if you curate a good one out off, say, 1000 random (midi) generations, you can compose a Hit with it

3- Performed : Videos are more difficult that images, but we are already conditionned to CGI + VFX heavy music videos (even low-prod ones). So I don't think it is difficult to generate an AI powered ("collage" style) deep-faked music video.

It is not over though. You did not mention if the sound is totally AI generated or if humains arrange/produce and play/sing in it.

As you can glimpse from the openAI jukebox [1], IMO it is convincingly generating these two elements with the three others into a coherent (less hallucinatory) whole that may be real difficult.

[1] https://jukebox.openai.com/

2 comments

A lot of innovation in music AI is currently happening at HarmonAI, a group within Stability AI. My guess is they will soon release some sort of "Jukebox 2.0" model.
> if you curate a good one

wow, didn't expect A&R to be the only part that couldn't be automated

There are already AI based Analytics Services used by labels/A&R. However, they analyse how tracks are already performing (= their emergence) to gauge the artists future success. Melody pre-analysis is not a big part of that (to my knowledge).

To return to my reply, I suppose you can use a second level AI to rate that first-level 1000 generated melodies, instead of manually curating THE ONE. I don't rule that out. Will it be more successful than a human-in-the-loop process? Maybe one day.