Where Is the AI Music? | Hacker News Reader

Where Is the AI Music?

33 pointsjay_kyburz32 comments

After seeing those TRON images this morning I've been playing with Midjourney all day. Version 4 really is amazing and I can see how fast and how much this tech will change the world.

Anybody know if there are AI music generators I can play with?

Would be cool to prompt with composer names, instruments, bands, moods, and feelings.

View on HN

One issue is the legality is much more risky than with visual art. While there's been talk of artists suing AI-art companies for using their art in training sets, and generating images "in their style", it isn't clear if there is a real legal case to be made. In music, artists have been suing each other for years over songs that sound superficially like each other even if the artist being sued didn't really intend to copy. I'd imagine AI music would run into similar problems.

I think sound is, perhaps counter-intuitively, harder than image.

We are much more forgiving of the poor framing of a Youtube video than traffic noise behind the narration in the same video.

Consider the difference in complexity between composition in painting against composition in music.

Or to put it another way, I suspect convincing visual AI tools are more accessible because it is an easier problem than convincing AI sound.

Good luck.

You are almost right. Your only mistake IMO is in confounding Music with "Music composition":

To use your expression, Music (and not sound) is indeed, "perhaps counter-intuitively, harder than image."

You also correctly stated the right reason why: "difference in complexity between the two types of compositions"

However, 40 Years of Electronic Music and Hip Hop music have utterly demonstrated how forgiving we can be of sounds like Cheesy synths and low-res samples loops, if they are culturally "organized".

This is good news for AI!

At least in the short term, the real creative potential of AI in music is in manipulating "Musical Sounds", not in generating "Musical Composition".

PS : I've recently made a short list of creative musical AI tools here https://rentry.co/Music-Creation-AI-Tool .

Missed an ‘s’ in your url https://rentry.co/music-creation-ai-tools

Images are only correlated on very small space/time scales. Audio/Music has correlation on multiple over-lapping time scales. We know this is true because Video filters are usually very small, bilinear, bi-cubic, operating on only a very small region. A large video filter doesn't make sense because there is no correlation. Audio requires attention to tone, transients, harmonics, voice envelopes, relationships between notes/scales, rhythym, recurring musical passages etc. etc. All these require filtering, or memory based processing over different time scales. So audio/music from a raw data perspective is much more difficult for an AI system compared to images. This is why image related AI is so prevalant, it is the low-hanging fruit due to small amount of correlation that the Neural Net needs to deal with.

I remember a youtube Video 3-5 years ago, where a dude explained the process really well and generated music e.g. from Mozart and it really sounded okay. It was not great, but this was 3-5 years ago. But i cant find his channel anymore :(

That is why i'm also confused how it didn't progress much further in the last 3-5 years, at least a little bit.

That sounds like the youtuber carykh.

Thanks! Couldn't find him anymore as it was a long time ago but i enjoyed his Videos back then.

DeepBach. This style of music is considerably easier to generate since timbre and dynamics are not generated.

Why aren't dynamics generated? Seems like a very important element to lose from the source material.

Dynamics aren't in the source material.

There's plenty of dynamic markings in music composed after the advent of the piano.

It would be in more modern music, but not Bach. Dynamics in music notation didn't come until the Renaissance.

Ah, I thought for Bach it was fair enough but Mozart is a different matter.

There is a huge difference between generating a series of notes, and generating actual audio. There can be dynamics in both. Just focusing on the actual notation is a much simpler problem. Trying to generate audio, requires not just notes, but tone harmonics, stochastic noise, rhythym, all sorts of properties that ears are extremely sensitive to.

No, the dynamics can be generated just like the notes. Actual scores have markings for dynamics and midi provides 1-127 dynamic points.

The dynamics are an important part of the music and can't be ignored

https://aiva.ai./

https://www.strofe.com/get-started

Algorithmic composition has been around for a long time now but most has not been very good.

There is also a huge difference in the quality of amateur photography vs the quality of amateur music composition. I have some of stuff on soundcloud that is absolutely terrible compared to any picture I can take with an iphone.

If you use midi data of the classical masters what is the point? Bach wrote over 1000 compositions himself and most people only know Cello Suite No. 1, if that.

OpenAI jukebox was pretty impressive at the time. As mentioned in this thread, the legality is probably where the issue lies, not the feasibility at this point.

https://soundraw.io/ has impressed me, and it's both a permissible license as well as reasonably affordable. I'd say only about 10-20% of what it makes is acceptable to me, but that fraction of it is actually pretty catchy.

Check this one, AI generated techno made using the channel HÖR Berlin as the training set: https://youtu.be/UPs5L-1Amo8

I've seen/heard it used to supplement existing compositions (a la NodeMusic. com) but not used unsupervised. As the site suggest, how will an AI know it's written good music?

there is this technical death metal (Archspire) generation from April 2019 [0]. The 24/7 stream is available. [1]

[0]: https://www.sciencealert.com/new-24-7-ai-generated-death-met...

[1]: https://youtu.be/MwtVkPKx3RA

I'm sure they are already using it to generate music for the masses. As exchange for using this secret software, the artists have to cover one eye in photoshootings.

once image models are good enough, you can train them to generate spectrograms images, and convert them back to audio to listen to the generated music.

This is one approach, but there are also models that operate in the time domain, like the one by Harmonai.

I work in Animation, and I am extremely interested in an AI that produces 3D meshes & materials.

Check out nerfs

Same here! It's happening, but there aren't any available to the public, at least with a GUI.

I've been using AI that generates height maps using monocular depth estimation, and modeling from there. Combining this with Midjourney gives pretty fun results.

I see it being used modularly. Example is using "AI" to automate post production mixing.

Commercial: aiva.ai, soundful.com Open source: jukebox

We are not there yet.

Reply on news.ycombinator.com