3 ms·

Note that the orchestration on the available AIVA tracks is (at least as far as one can tell) entirely manual. Even the music itself may well be more like the r

Note that the orchestration on the available AIVA tracks is (at least as far as one can tell) entirely manual. Even the music itself may well be more like the result of a process involving both the 'AI' and the human producer, than something truly generated 'end to end',[1] free of manual intervention! (Not that there's anything wrong with the former; it might be the best way of ending up with a truly compelling and impressive end-product!) (I don't quite recall off-hand if it was these folks or someone else who claimed to have fed their AI a corpus of tens of thousands of classical pieces in "digitized sheet music" form as part of the training process - something that would be a treasure trove for academic music corpus studies, if it existed!) [1] End-to-end systems do exist, and their output is quite interesting, but in a "quirky, weird but it actually sounds nice" way - sorts of like the musical analog of the DeepDream doggie pictures. Arguably the most successful (of the ones that are openly available) is "Hexahedria"/Daniel Johnson's "biaxial music composition" (aka "tied parallel networks") project, which was featured here before. The "BachBot" and "DeepBach" projects also deserve mention, but their scope is rather limited as their names suggest.

byproxy 8y ago
I'm sure you know about it, but IMSLP (https://imslp.org/ https://imslp.org/) has quite a few scores from which they could pull.
- zozbot123 8y ago
  These are not "digitized" in the relevant sense, though. The OpenScore project (https://www.openscore.cc https://www.openscore.cc ) makes it quite clear what is meant by that. And it's hard work, though in principle some AI engine of a very different sort might be able to help with it.
  - byproxy 8y ago
    Ah, I see. So, some kind of OCR for sheet music, then?
inetsee 8y ago
I am very interested in the "freestyle" model of chess, where one or more human players are paired with one or more chess engines to form a team, playing against other teams. This model has demonstrated that these teams can play stronger chess than either the human players or the chess engines. I believe that this same model can be applied to many more creative applications, including music. A lot of the pure AI creative applications I've seen still have a long way to go before they get close to what talented human creatives can do.
pierreb-aiva 8y ago
Hi there! :) Indeed, up till now, all of the orchestral tracks we released where orchestrated by humans. However, we've just finished automating that part of the process as well; our Beta will allow users to generate scores that are orchestrated for a wide variety of ensembles. And regarding our database, we have over 30.000 scores right now, digitized in MIDI format. :)

byproxy 8y ago

I'm sure you know about it, but IMSLP (https://imslp.org/ https://imslp.org/) has quite a few scores from which they could pull.

zozbot123 8y ago
These are not "digitized" in the relevant sense, though. The OpenScore project (https://www.openscore.cc https://www.openscore.cc ) makes it quite clear what is meant by that. And it's hard work, though in principle some AI engine of a very different sort might be able to help with it.
- byproxy 8y ago
  Ah, I see. So, some kind of OCR for sheet music, then?

inetsee 8y ago

I am very interested in the "freestyle" model of chess, where one or more human players are paired with one or more chess engines to form a team, playing against other teams. This model has demonstrated that these teams can play stronger chess than either the human players or the chess engines. I believe that this same model can be applied to many more creative applications, including music. A lot of the pure AI creative applications I've seen still have a long way to go before they get close to what talented human creatives can do.

pierreb-aiva 8y ago

Hi there! :) Indeed, up till now, all of the orchestral tracks we released where orchestrated by humans. However, we've just finished automating that part of the process as well; our Beta will allow users to generate scores that are orchestrated for a wide variety of ensembles. And regarding our database, we have over 30.000 scores right now, digitized in MIDI format. :)