5 ms·

Generate Musical Accompaniment with R

97 points by hackoo 5y ago

CornCobs 5y ago
Cool! Machine learning analysis of music seems to largely be at the audio or midi level, and not really at the symbolic level. What kind of learning techniques would be applicable to structures like musical meter?
- zeckalpha 5y ago
  MIDI is semi-symbolic.
  - TheOtherHobbes 5y ago
    MIDI happens at the note level, and the controller level for parameter management. MIDI has no concept of chords or chord languages.
  - CornCobs 5y ago
    It's still on the physical time scale and presents music as a linear sequence. I believe symbolic music representation is more of a graph than a list
adamnemecek 5y ago
I’ve been working on and IDE for music composition. Launching soon https://ngrid.io https://ngrid.io.
- billfruit 5y ago
  Will it have a tracker interface?
  - adamnemecek 5y ago
    No, an advanced piano roll.
setgree 5y ago
This looks like a cool project and I look forward to trying it later. General comment: one reason that the R package ecosystem often feels creaky and brittle, I think, is that a lot of packages (like this one) are pretty far afield from R’s original intended uses of statistics and visualization. Here, we see the use of integers to denote pitches strung together as a vector to generate a sequence. That’s not precisely a misuse, but it’s definitely misaligned with the original vision, and trying to build things on top of that misaligned infrastructure is a potential vector of packages’ breaking and general irreproducibility.
- adeelk93 5y ago
  R’s package ecosystem is creaky and brittle even for statistics and visualization. I don’t think “misaligned with the original vision” is really relevant. Like, it has nothing to do with someone making a package to work with music in R. Not sure how any of this is really relevant here. This is a general R gripe.
  - malshe 5y ago
    > R’s package ecosystem is creaky and brittle even for statistics and visualization Explain what you mean by this. It’s a big claim without any supporting evidence.
    - adeelk93 5y ago
      Sure, I thought this was a common experience, and I’d love to hear how you get around it. The tidyverse is absolutely full of breaking changes. Two big ones off of the top of my head: dropping underscore functions like mutate_, and turning common column names into new function names. So in different organizations you might see two patterns: * Monolith - everyone uses a set of R packages from 2017 or whatever (frozen in time either through MRAN or Nexus) * Wild dockerization - one team’s internal tooling is incompatible with the other team’s, so working across teams you gotta rely on Docker a lot Why? Cause there’s not really a good way to pin versions besides shrinkwrapping, at least in my experience. We tried Drake, but Drake itself had so many breaking changes that it gained a really bad rep (from what I hear from colleagues, haven’t used it myself). Tried renv but it was destroying some of our analyst’s setups (like, you just get a bomb icon in Rstudio). Didn’t troubleshoot it too much, kinda just gave up. I don’t like forcing a specific vintage of packages, but I like everything else less.
sidedishes 5y ago
I've been doing one-hour time limit song writing challenges and in those I need to write accompaniment fast. I gotta say, this generation is a really close reduction of what I actually do -- come up with some accompaniment figure over a chord or two, then transpose it and tweak chord tones for the next measure. Maybe add a part or do something different if it's a section change. IMO this post is a great example of how to mix variation and repetition to balance interest and familiarity. Even if the generation itself has limitations it seems like a great springboard for manual touches like dissonance tweaking, or new sections, or fills / transitions -- arguably the cool creative part after all anyway! There's a lot of repetition in music, and when writing, it's hard to decouple it from points of interest. A workflow that enables such is very exciting. (Taking this opportunity to share my multiplayer piano roll editor: https://yuxshao.github.io/ptcollab/ https://yuxshao.github.io/ptcollab/ )
bambax 5y ago
Isn't this fundamentally describing an arpeggiator? Using Reaper I made a simple arpeggiator extension: it takes input from one track for the chords and from another track for the notes, and it maps the notes from the notes track to the available notes from the chords track. It's like a "snap to scale" but it snaps to the current chord. It's much better than a normal arpeggiator plugin IMHO because you don't have to program patterns in the UI of a plugin; you can do anything in the notes track and it will always work. The limitation, as with any arpeggiator, is that it can't play notes outside of the current chord (but it's possible to have "chords" with many notes on the chords track to allow more freedom). One way that I use it is I write a chord progression, then I play over it using the extension, and then I adjust the recording by moving notes to more interesting positions. This song was made this way: https://open.spotify.com/track/5TxVfIf9JUAhCEL3O5cWXT https://open.spotify.com/track/5TxVfIf9JUAhCEL3O5cWXT (It's not perfect and the sound is kind of bad, but it's a start).
- motohagiography 5y ago
  Edit: your stuff is beautiful btw. Less than an arpegiator, I think it's closer to 1/f or fractal iteration. I've spent the last few days working out a similar hypothesis using a "fractal sequencer" module for my eurorack rig. (https://www.qubitelectronix.com/shop/bloom https://www.qubitelectronix.com/shop/bloom) Essential idea is it takes your sequence of notes, creates a bunch of branches that are essentially folds, and then iterates through them. The track linked is two sequences iterating the 1st 3rd 6th 8th of a scale across two different oscillators creating a kind of bleep-bloopy counterpoint in the background, and then adding a manual three note lead with some drum breaks and drops to tie it together. I wanted to see how different fractal randomness sounded from intentional complexity. (https://soundcloud.com/n-gram-music/beatrice https://soundcloud.com/n-gram-music/beatrice) It's still sounds random'ish because it plods, where a real composer would add intention, emphasis, and ornamentation, but it's still within the domain of clearly musical. I'd say it's based on minimalism and Arvo Part's tintinabuli method, but I think one should be an actual serious musician before making lofty claims like that. (I am not) However, this idea of whether there is a big difference between 1/f fractal complexity in a consonant set of notes, and actual human intent is interesting. I suspect Bach (and before him, Vivaldi) had an underlying system that was based on fractal iterations of each theme, which is what sent me down this (rather nutty) path, but the journey continues to amuse.
pvaldes 5y ago
system('vlc Lorde-Melodrama.avi')

97 points by hackoo 5y ago

Cool! Machine learning analysis of music seems to largely be at the audio or midi level, and not really at the symbolic level. What kind of learning techniques would be applicable to structures like musical meter?

zeckalpha 5y ago
MIDI is semi-symbolic.
- TheOtherHobbes 5y ago
  MIDI happens at the note level, and the controller level for parameter management. MIDI has no concept of chords or chord languages.
- CornCobs 5y ago
  It's still on the physical time scale and presents music as a linear sequence. I believe symbolic music representation is more of a graph than a list

adamnemecek 5y ago

I’ve been working on and IDE for music composition. Launching soon https://ngrid.io https://ngrid.io.

billfruit 5y ago
Will it have a tracker interface?
- adamnemecek 5y ago
  No, an advanced piano roll.

setgree 5y ago

This looks like a cool project and I look forward to trying it later. General comment: one reason that the R package ecosystem often feels creaky and brittle, I think, is that a lot of packages (like this one) are pretty far afield from R’s original intended uses of statistics and visualization. Here, we see the use of integers to denote pitches strung together as a vector to generate a sequence. That’s not precisely a misuse, but it’s definitely misaligned with the original vision, and trying to build things on top of that misaligned infrastructure is a potential vector of packages’ breaking and general irreproducibility.

adeelk93 5y ago
R’s package ecosystem is creaky and brittle even for statistics and visualization. I don’t think “misaligned with the original vision” is really relevant. Like, it has nothing to do with someone making a package to work with music in R. Not sure how any of this is really relevant here. This is a general R gripe.
- malshe 5y ago
  > R’s package ecosystem is creaky and brittle even for statistics and visualization Explain what you mean by this. It’s a big claim without any supporting evidence.
  - adeelk93 5y ago
    Sure, I thought this was a common experience, and I’d love to hear how you get around it. The tidyverse is absolutely full of breaking changes. Two big ones off of the top of my head: dropping underscore functions like mutate_, and turning common column names into new function names. So in different organizations you might see two patterns: * Monolith - everyone uses a set of R packages from 2017 or whatever (frozen in time either through MRAN or Nexus) * Wild dockerization - one team’s internal tooling is incompatible with the other team’s, so working across teams you gotta rely on Docker a lot Why? Cause there’s not really a good way to pin versions besides shrinkwrapping, at least in my experience. We tried Drake, but Drake itself had so many breaking changes that it gained a really bad rep (from what I hear from colleagues, haven’t used it myself). Tried renv but it was destroying some of our analyst’s setups (like, you just get a bomb icon in Rstudio). Didn’t troubleshoot it too much, kinda just gave up. I don’t like forcing a specific vintage of packages, but I like everything else less.

sidedishes 5y ago

I've been doing one-hour time limit song writing challenges and in those I need to write accompaniment fast. I gotta say, this generation is a really close reduction of what I actually do -- come up with some accompaniment figure over a chord or two, then transpose it and tweak chord tones for the next measure. Maybe add a part or do something different if it's a section change. IMO this post is a great example of how to mix variation and repetition to balance interest and familiarity. Even if the generation itself has limitations it seems like a great springboard for manual touches like dissonance tweaking, or new sections, or fills / transitions -- arguably the cool creative part after all anyway! There's a lot of repetition in music, and when writing, it's hard to decouple it from points of interest. A workflow that enables such is very exciting. (Taking this opportunity to share my multiplayer piano roll editor: https://yuxshao.github.io/ptcollab/ https://yuxshao.github.io/ptcollab/ )

bambax 5y ago

Isn't this fundamentally describing an arpeggiator? Using Reaper I made a simple arpeggiator extension: it takes input from one track for the chords and from another track for the notes, and it maps the notes from the notes track to the available notes from the chords track. It's like a "snap to scale" but it snaps to the current chord. It's much better than a normal arpeggiator plugin IMHO because you don't have to program patterns in the UI of a plugin; you can do anything in the notes track and it will always work. The limitation, as with any arpeggiator, is that it can't play notes outside of the current chord (but it's possible to have "chords" with many notes on the chords track to allow more freedom). One way that I use it is I write a chord progression, then I play over it using the extension, and then I adjust the recording by moving notes to more interesting positions. This song was made this way: https://open.spotify.com/track/5TxVfIf9JUAhCEL3O5cWXT https://open.spotify.com/track/5TxVfIf9JUAhCEL3O5cWXT (It's not perfect and the sound is kind of bad, but it's a start).

motohagiography 5y ago
Edit: your stuff is beautiful btw. Less than an arpegiator, I think it's closer to 1/f or fractal iteration. I've spent the last few days working out a similar hypothesis using a "fractal sequencer" module for my eurorack rig. (https://www.qubitelectronix.com/shop/bloom https://www.qubitelectronix.com/shop/bloom) Essential idea is it takes your sequence of notes, creates a bunch of branches that are essentially folds, and then iterates through them. The track linked is two sequences iterating the 1st 3rd 6th 8th of a scale across two different oscillators creating a kind of bleep-bloopy counterpoint in the background, and then adding a manual three note lead with some drum breaks and drops to tie it together. I wanted to see how different fractal randomness sounded from intentional complexity. (https://soundcloud.com/n-gram-music/beatrice https://soundcloud.com/n-gram-music/beatrice) It's still sounds random'ish because it plods, where a real composer would add intention, emphasis, and ornamentation, but it's still within the domain of clearly musical. I'd say it's based on minimalism and Arvo Part's tintinabuli method, but I think one should be an actual serious musician before making lofty claims like that. (I am not) However, this idea of whether there is a big difference between 1/f fractal complexity in a consonant set of notes, and actual human intent is interesting. I suspect Bach (and before him, Vivaldi) had an underlying system that was based on fractal iterations of each theme, which is what sent me down this (rather nutty) path, but the journey continues to amuse.

pvaldes 5y ago

system('vlc Lorde-Melodrama.avi')