Neural networks and machine learning for music applications

Programming applications for making music on Linux.

Moderators: MattKingUSA, khz

Post Reply
Basslint
Established Member
Posts: 1511
Joined: Sun Jan 27, 2019 2:25 pm
Location: Italy
Has thanked: 382 times
Been thanked: 298 times

Neural networks and machine learning for music applications

Post by Basslint »

What areas of music-making do you think can be improved via neural networks and machine learning?

I think music transcription is one of those areas in which supervised learning could help a lot, better if done on a single instrument with a dry signal. For example, we can play thousands of guitar chords, label them properly, and give them to the algorithm, which should be able to transcribe guitar at the end. Then with a little bit of fine tuning (adapt the model to harmony theory, since the same chord can have a different name in a different context) we can release that trained model under a FLOSS license and enjoy getting better at jazz improvisation :lol: What do you think about it?
The community of believers was of one heart and mind, and no one claimed that any of his possessions was his own, but they had everything in common. [Acts 4:32]

Please donate time (even bug reports) or money to libre software 🎁

Jam on openSUSE + GeekosDAW!
CrocoDuck
Established Member
Posts: 1133
Joined: Sat May 05, 2012 6:12 pm
Been thanked: 17 times

Re: Neural networks and machine learning for music applications

Post by CrocoDuck »

Maybe mixing and mastering? Not to do all the job, but to automatize all the most tedious tasks. Like: "AI, please set the project up when I brew my tea, then I will be back tacking care of the details".

It would also be cool to "describe" to a plugin what you are after, like: "AI, I want a warm fuzz-like distortion, can you brew a prototype for me? Then I will take care of the details on it".

Composition also. You know, when we compose, as humans, we often do not realize how narrow minded we are. I am from the west and I cannot for the life of me think outside the confines of 12 tones equal temperament. You can give any temperament and any whacky time subdivision scheme to an AI and be like "compose something sad". Not sure how you would train that AI, but assuming you can than it would come up with something. Maybe absurd, but definitely not something you would think about. Maybe one section of it can inspire you, and you can expand upon it. It is as close as composing together with aliens as it gets. My guitar teacher used to get Band in a Box solo generator to create a solo. They are whacky and sound like cr$p, but sometimes they get this semi-random thing that makes you go "I like that, I would have never though of making a phrase like that, I am gonna use that!".

You know, I like the idea of using AI as some kind of help. Take care of the mundane stuff, let me focus on the things I care most about. I think this is the more "within reach" use case for it. But I think it can also serve to inspire us. You could train AI with all kinds of music humans invented, it can smash them together and generate things nobody could have though about, from which you can draw inspiration and build upon.

As a futuristic note, the thing that frustrates me the most about music making is that I have music in my head, and I cannot in any way turn it into reality so that it sounds as it does in my head. There is always a gap and that frustrates the living hell out of me. One of my favorite sci-fi writers, Alastair Reynolds, in one of his books describes this girl that is a sculptor. She has implants in her brain and all she has are ideas, and she can visualize her art in her brain with high detail. The implants take it from there and drive robots that make the sculptures exactly as she intended. Not sure when, or even if, down the line that is, but geez I would love it. My creativity wired straight to the end result, no fuss, no confounding factors, no jitter of my nervous system, no unconscious bias when I turn knobs on a GUI. The sound as I want it, for real. I would love that: to have a direct interface from ideas, deep down me, to reality.
Last edited by CrocoDuck on Sun Mar 08, 2020 2:49 pm, edited 1 time in total.
User avatar
SpotlightKid
Established Member
Posts: 250
Joined: Sun Jul 02, 2017 1:24 pm
Has thanked: 48 times
Been thanked: 54 times

Re: Neural networks and machine learning for music applications

Post by SpotlightKid »

* Analizing a piece wrt to musical theory given sheet music or a (possibly quantised) MIDI file.
* Generating notation from a MIDI file of a real performance, i.e. dealing with timing and tempo variations and expression (pitch bend, modulation, expression pedal etc.)
* We'll probably see human voice modelling rather sooner than later, i.e. grafting the tonal and performance characteristics of one vocal performance onto another.
Basslint
Established Member
Posts: 1511
Joined: Sun Jan 27, 2019 2:25 pm
Location: Italy
Has thanked: 382 times
Been thanked: 298 times

Re: Neural networks and machine learning for music applications

Post by Basslint »

https://storage.googleapis.com/ddsp/index.html#modular

Just saw this article and I thought I'd share it here, as additive resynthesis is very interesting. I think that in the future, it might replace soundfonts entirely.

Edit: it's libre (Apache 2.0) software by the way https://github.com/magenta/ddsp
The community of believers was of one heart and mind, and no one claimed that any of his possessions was his own, but they had everything in common. [Acts 4:32]

Please donate time (even bug reports) or money to libre software 🎁

Jam on openSUSE + GeekosDAW!
diegorapoport
Posts: 1
Joined: Tue Jul 21, 2020 4:06 pm

Re: Neural networks and machine learning for music applications

Post by diegorapoport »

I have an idea of a project to make a melody recognizer, or sort of. I'm already aware of some trouble but it would be some interesting challenge to implement it. If anyone is interested, please send me a message or email.
Post Reply