Anyone else using ChatGPT to make music? (2023) | Hacker News Reader

Anyone else using ChatGPT to make music? (2023)

community.openai.com

29 pointswslh27 comments

Indirectly. I used ChatGPT to help make this tool, https://editor.p5js.org/spDuchamp/full/zgtkE2xik (ChatGPT+p5js is a handy way to work through an idea.)

That was used to design a balanced irrational arrangement of magnets for a rotary magnetic bow. https://twitter.com/CommonStef/status/1654663207707979776

I use that rotary magnetic bow to do experiments and make music. https://soundcloud.com/spduchamp/sets/6-vs-5-magnet-rotary-m...

Not chat gpt specifically. My mini project to get fammilar with these models was, taking a bunch of guitar tab, make a tokenizer for tab, trained a small gpt2 like network and had it write new guitar tab.

Some of it was alright. It was cool seeing it have musical structure. I think I need more gpus and faster ones to get it to do anything really interesting though.

I wanted to try fine tuning it on specific bands and artists after training to see if it could copy their style lol.

I was wondering how in the world they got ChatGPT to do this, but it's actually interesting:

  "Got ChatGPT to write some midi music ... used another ChatGPT prompt to create the computer generated voice in python and compiled the track." 

  "ChatGPT told me to write a MusicXML file that I uploaded to https://www.sinsy.jp/ ... Sinsy is an HMM/DNN-based singing voice synthesis system. You can generate a singing voice sample by uploading the musical score (MusicXML) to this website."

WAV files for both attached in the original post.

Btw there is a lot of AI that creates "audio" - voiceovers, background music, etc. So you don't need to use ChatGPT for this, but still interesting it can produce MIDI and MusicXML files from it.

Check out suni.ai, especially outputs from the new v3 model released a couple days back. The quality seems to have gone way beyond background music. E.g https://app.suno.ai/song/efc4e137-3464-49dc-9409-445b9943e3b...

OK this is getting fucking insane, wow!

https://app.suno.ai/song/d6afe7c6-35be-4fde-9d83-d0884cf0a38...

Even pre-AI, a lot of thought had gone into the notion of defining music parametrically. Example - https://www.researchgate.net/publication/329281114_Thinking_...

It will be interesting to see how parametric definitions can be augmented by the strengths of LLM.

I use ChatGPT to help me understand and use popular music software as well as some music theory questions, but I don't use it to generate actual melodies. I use it in a similar way for creative writing - analysis, synopses, ideas, but not for generating actual narrative text.

There are models that generate music directly.

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

Musiclm is one.

WavTool does this! https://wavtool.com/

They've got an AI "composer" that can make midi songs that really can sound great!

I have wondered what grooves it could come with using https://sonic-pi.net/

I've seen countless advertisements, mainly on Facebook, promoting AI-powered VSTs. As anyone here used these? Are they any good?

Anything even remotely interesting advertised on Facebook is a scam.

Taking a broad view of music, ChatGPT has been great at creating lyrics for songs my DND character sings.

Sorry, how can you output MIDI from LLM models?

beep boop boop beep - Mr. Krabs

"Brown-chicken-brown-cow": the '70s

https://www.musicxml.com

There are this and other methods discussed in the linked thread.

Why would I possibly do that? I would no more do this than use an AI to listen to music for me. What's next I get an AI to have sex for me, to watch the sunset so I can focus on work?

AI should do tedious stuff for artistic people. Instead it’s doing artistic stuff for tedious people.

As a non-artist sometimes I find artists somewhat tedious, they all seem to pretend they are doing something unique and creative yet it doesn’t look particularly interesting or different to me, instead more tedious.

You’re a creative professional with deep domain expertise, yes? If you don’t mind sharing what you do, I think I can make the point in a way that will resonate more for you.

I’m a software engineer and also a quantitative researcher/analyst.

I think you are coming at it from a different direction than the artist. You want them to create something for you to consume, and the value is whether you enjoy it. The artist is scratching their personal itch, and the value is them enjoying the creation process.

In that case AI wouldn’t bother any of them then.

Reminds me of that viral piece a year ago about a guy who wrote a book for his daughter. And by "write a book" I mean he used ChatGPT to come up with the story, Stable Diffusion to generate the pictures, and a vanity press to print it.

https://youtu.be/QKZNOnLUBmQ

Depends on what you mean by music. Devices have long been able to make images ... but can they make 'art'? Not trying to be 'argumenty' ... I can (and have) fed (random, as well as 'generated') ( patterns of ) numbers into a sequencer, hooked it up to a synth, and it produced tones). Just like Conlon Nancarrow did.

I thought it was sometimes 'music-ey'. (Often depending on the patch.) My partners-in-crime didn't all agree. I'm just wondering about when it's 'music'.

It's one thing to 'land' on the Moon and another to 'land' upright. ;-}

'Won't Get Fooled Again': https://www.youtube.com/watch?v=O5voNyRmvXs

Art by definition requires self-expression. There is no self to express = there is no art.

A generated piece of music or visual work is art in the same way a tree or the Moon is art: not at all. It may seem pretty to a human observer, but that’s orthogonal to it being art, it’s just an artifact that exists.

Unfortunately, the wide availability of art-mimicking non-art artifacts devalues art.

However, paradoxically, I suspect the wide availability of ML-powered generative art is also going to raise the value of direct improvisational experience, so in the end it’s a mixed bag.

First we had no way to record, so we gathered to listen to music performers display live their hard-earned skill.

Then we got a way to record, so we could hear live performers display their skill at any time.

Then we got increasingly sophisticated audio recording, mixing, and reproduction tools, which meant what we hear on the record is not entirely performer’s mastery, but also the mastery of audio engineers & producers.

Now we cannot trust that what we listen to is anyone’s mastery, really. It can be algorithmically composed, algorithmically performed, and there is no specific set of individuals whose hard-earned skill we could admire.

The only way to experience that truly would either be to listen to pre-2024 works, or to attend live performances.

Reply on news.ycombinator.com