3 ms·

Having worked on procedural music for some time, I really can't see how simple Markov-style approaches like this are likely to have interesting results. You can

Having worked on procedural music for some time, I really can't see how simple Markov-style approaches like this are likely to have interesting results. You can certainly create sequences of notes that have statistical similarities to the notes in a training set, but we don't hear statistical similarities, we hear phrases and refrains and call-and-responses and whatnot - higher-order structures that markov models seem really ill-suited to. It strikes me as analogous to generating text from an ML model trained on a bunch of short stories. A short snippet of the results may sound convincing, but obviously a whole story generated that way would be gibberish. I think the same sort of thing is happening in the example songs - any given chord change or three-note phrase may sound musical, but over the course of several measures it's not distinguishable from random notes.

coldtea 8y ago
>You can certainly create sequences of notes that have statistical similarities to the notes in a training set, but we don't hear statistical similarities, we hear phrases and refrains and call-and-responses and whatnot - higher-order structures that markov models seem really ill-suited to. You can have markov models at various levels of granularity...
- TheOtherHobbes 8y ago
  You can, but OP didn't. And you need to understand what you're Markovising - which apparently OP didn't either.
  - AstralStorm 8y ago
    OP built a whole song self similarity matrix and regularised on that, which forces the model to adapt this kind of high level structure.
    - TheOtherHobbes 8y ago
      Neither transition tables nor naive self-similarity searches are an effective way to model this domain.
- fenomas 8y ago
  > You can have markov models at various levels of granularity... This is neither here nor there. The article already applies models to two separate levels of granularity (chord progressions, and melody notes over a given chord). But for higher order structures the hard part is decomposing source music into data points that typify its structure - if you did that then applying a markov (or other) model to the data would be straightforward.
AstralStorm 8y ago
Don't worry, RNN generated version doesn't even generate a snippet of music. Calling it music is being generous, at best it is noodling. On the other hand, Markovian Models can segment and learn grammars. (Specifically probabilistic context free grammars.) The problem with these is overfitting.
abledon 8y ago
What’s the best github repo to get started with wav file data generated music... how big of gpu mem too? 6?11?24?gb
vidarh 8y ago
I think it might, but not without much more complexity. As you point out we hear phrases and refrains etc. - you'd need to both generate phrases, and then generate structure by making use of and transforming (mutating, transposing, investing etc.) phrases. Listening to his examples what struck me was exactly that: You could hear the outline of something, but it was killed instantly by the lack of internal consistency: What at first sounded like the outline of phrases were self similar, sure, but there was way too little outright repetition. It also fell flat at capturing well established rules of chord progressions - I have the same impression as you: individual changes were ok, but even with my very, very superficial understanding of composition I know that certain progressions add tension and others release it, and, and pairs of chords that work well together sound just plain weird if they're not part of a bigger structure that also follows the rules. I'm reminded of a program printed in Compute! or similar in the 80's that was meant to let anyone play music. It worked surprisingly well. The way it worked was simply to only allow note progressions that were likely to result in something that at least sounded "like music". But you still needed to understand rules for composition to go from something "like music" to something that sounded pleasant. I feel like his example falls in that category: He's made it sound "like music", but it's something most people can beat by buying a book on improvisation that explains a few simple chord progressions and some very basic rules. Going from that to something that sounds like an actual proper piece of music is another matter.

coldtea 8y ago

>You can certainly create sequences of notes that have statistical similarities to the notes in a training set, but we don't hear statistical similarities, we hear phrases and refrains and call-and-responses and whatnot - higher-order structures that markov models seem really ill-suited to. You can have markov models at various levels of granularity...

TheOtherHobbes 8y ago
You can, but OP didn't. And you need to understand what you're Markovising - which apparently OP didn't either.
- AstralStorm 8y ago
  OP built a whole song self similarity matrix and regularised on that, which forces the model to adapt this kind of high level structure.
  - TheOtherHobbes 8y ago
    Neither transition tables nor naive self-similarity searches are an effective way to model this domain.
fenomas 8y ago
> You can have markov models at various levels of granularity... This is neither here nor there. The article already applies models to two separate levels of granularity (chord progressions, and melody notes over a given chord). But for higher order structures the hard part is decomposing source music into data points that typify its structure - if you did that then applying a markov (or other) model to the data would be straightforward.

AstralStorm 8y ago

Don't worry, RNN generated version doesn't even generate a snippet of music. Calling it music is being generous, at best it is noodling. On the other hand, Markovian Models can segment and learn grammars. (Specifically probabilistic context free grammars.) The problem with these is overfitting.

abledon 8y ago

What’s the best github repo to get started with wav file data generated music... how big of gpu mem too? 6?11?24?gb

vidarh 8y ago

I think it might, but not without much more complexity. As you point out we hear phrases and refrains etc. - you'd need to both generate phrases, and then generate structure by making use of and transforming (mutating, transposing, investing etc.) phrases. Listening to his examples what struck me was exactly that: You could hear the outline of something, but it was killed instantly by the lack of internal consistency: What at first sounded like the outline of phrases were self similar, sure, but there was way too little outright repetition. It also fell flat at capturing well established rules of chord progressions - I have the same impression as you: individual changes were ok, but even with my very, very superficial understanding of composition I know that certain progressions add tension and others release it, and, and pairs of chords that work well together sound just plain weird if they're not part of a bigger structure that also follows the rules. I'm reminded of a program printed in Compute! or similar in the 80's that was meant to let anyone play music. It worked surprisingly well. The way it worked was simply to only allow note progressions that were likely to result in something that at least sounded "like music". But you still needed to understand rules for composition to go from something "like music" to something that sounded pleasant. I feel like his example falls in that category: He's made it sound "like music", but it's something most people can beat by buying a book on improvisation that explains a few simple chord progressions and some very basic rules. Going from that to something that sounds like an actual proper piece of music is another matter.