Hacker News Reader

Comment by caseyross | Hacker News Reader

With the success of AI content generation for text (GPT-3 etc.) and images (DALL-E etc.), it seems inevitable that music will soon be targeted as well, if not already.

What's particularly interesting in the music sphere is that there are already well-established trends towards building a sort of ambient, atmospheric, generated soundscape. (For example, the famous "Lo-Fi girl" stream.) AI-generated content is a very natural progression here.

Regarding the broader pop music industry, a "GPT-3 for music" would likely further inequalize the relative power of labels and musicians. If people who control music distribution can easily make hit songs without needing to hire songwriters, arrangers, or performers, they surely will do so. I can imagine a lot of music-related occupations potentially having to pivot to rely much more heavily on live performances to make any money.

View on HN

They always have relied on live performance (and to a greater extent, education). Talk to some professionals in the industry some time, they're not going to talk about records as a source of income.

I could write a ten page rant on why AI and automation aren't a threat to musicians but it basically boils down to the fact that music is a human spectacle and we will continue to grow the industry through performing live music much like a robot that can mimic Tom Brady isn't a threat to teams selling tickets to see Tom Brady play live and people to watch it on TV (which exists, by the way).

At the end of the day, technology that lowers the barrier to entry for records and distribution is a massive boon to the industry at all levels. The rising tide lifts all boats. The companies doing interesting things with AI in music aren't just generating old shit, they're making tools to give to the next generation of creators to create new shit that no training can replicate, because it has never been done.

Your comment ignores 99% of genres.

I don't think it does, if you are alluding to electronic music that's also just as much about the human aspect of it. Lets be real, at the commercial end of the spectrum electronic music is trivial to make, but no one wants to hear "Song seed 2aslk3j25lh" they want to hear what their "Music Hero" has made. Another aspect is DJing, DJs don't even pretend to have made the music yet people will come out just to hear a specific DJ play other peoples songs. It's almost entirely about the human figurehead, the popularity contest, the status and the fashion of it all.

Some labels might try to present artists that have AI music, and no doubt it would still be consumed. But it's a huge risk when it to the human aspect of it as people want authenticity.

Music is consumed in a huge array of contexts though, where I think AI music will end up is as music in movies, backing tracks in adverts and youtube videos, as filler music for all kinds of other media.

Let's be realistic here. Hatsune Miku was created before 2010. While I dislike the genre, they've completely sold out all tickets to these festivals where "she" "performed".

You're really not thinking it through if you think any genre isn't going to fundamentally change once the industry starts to push virtual artists. They can be perfect and relatable to teenagers. You really don't need physical people to pull off a good festival, a well orchestrated 3d avatar is likely even better because they can be bigger and seen from the back

Hatsune Miku kind of helps my argument here, because they still needed a personified identity for the music even though none of it's real. So then we ask, if Hatsune Miku could be manifested into the real world, do you not think their fans would be absolutely ecstatic? Behold, humans, the solution to that problem.

I know it's possible to get some people excited about an avatar, but I'm going to argue that the vast majority of people would prefer humans for as long as they can get them.

You might want to check out a Hatsune Miku concert on YouTube because humans are actually entirely redundant if you can just generate the song. (You'll probably want to mute the audio though, otherwise your ears might start to bleed)

There is a 3d avatar dancing on the floor. The holographic technology in the context of concerts is incredible at this point.

As a matter of fact, its likely going to be in favour of AI if you consider VR headset etc, as the coming generations will be able to interact with the virtual artists, giving the producer am even easier time to get money from the consumers.

I'm not looking forward to that future to be honest.

Example: https://youtu.be/PlQIdq5mv_k

/Edit: After thinking about it some more: I think I agree that the potential music generation isn't going to change anything by itself. It's just another building block that will enable the music industry to eventually remove real humans from the equation. while we're slowly progressing on that path, the music generation alone won't push us to the logical conclusion.

In Orwell's "1984", a machine called the Versificator generated the music and literature for consumption by the proles.

Yes, it's inevitable. Some parts of this process are harder than than others, though. I think I've handled the hardest part, which is making catchy melodies (https://www.youtube.com/playlist?list=PLoCzMRqh5SkFwkumE578Y...). Live music is already over 50% of music industry revenue, so the transition could be a bit easier for musicians than for visual artists.

Didn't find the melodies catchy. Sorry to tell you that.

In fairness, I'd bet that most of your favorite tunes would become very un-catchy if they were stripped of harmony and percussion, and played by a quantized, expressionless square wave

The melody style reminds me of a songbook I have of Tin Pan Alley songs.

Generated music has been around for at least several years, for example https://generative.fm/ (no affiliation).

AI that imitates well-known composers goes back much further, I remember attending a concert of "virtual mozart" works at UC Santa Cruz back in the late 80s/early 90s

https://www.theguardian.com/technology/2010/jul/11/david-cop...

OpenAI (the people behind GPT-3) already make one. It's called JukeBox. It can even make new unique songs in the style of existing artists with a simulation of their vocal style.

https://openai.com/blog/jukebox/

Considering that significant portion of the pop music produced in the last 30 years is made by a few people like Max Martin, one can assume that there's a formula for writing music that people love. Seems promising for AI to be honest.

Generated music but with dance troupes? Ala K-Pop?

That's the plot of Norman Spinrad's book Little heroes. It starts with the idea that Music record companies can only make bad automated music, and in order to make a hit, they hire real musicians.

It's a great read! https://www.goodreads.com/book/show/719900.Little_Heroes?ref...

I think music, especially pop music, is in a completely different ball park than text or images. An AI can't just make a catchy song only with a training model and some rules. Successful musicians have a feel for what we will like and what will be trendy. An AI can't do that, at least not for a while to come. Well, thats just my thoughts, I'm not an expert on this.

Music can be algorithmically generated since it's formulaic, what I think machine learning models will struggle with at first is maintaining those formulas throughout a song. I imagine AI produced songs will/do sound winding and unhinged at first. All the right ingredients but stewed up in a pot.

Like Dall-E. What you'll get is music which apes music tropes really well, but strangely intentionless and disassociated. For genres like ambient where the intention is to be intentionless and disassociated (or can be), the result is like 'oh hey, AI is here!'.

It's weird. You could extrude endless amounts of The Caretaker, "Everywhere At The End Of Time" if you specified the recipe, with a 'deterioration dial' and some coding to determine how you vary the output. But you'd be free-riding on a pre-existing artistic intention that was originally implemented in dramatic, bold manner. To extrude endless amounts of this stuff is both fairly trivial and missing the point completely…

Reply on news.ycombinator.com