< Back to IRCAM Forum

Music and Machine Learning: the Issue of Euphony

General Forum Workshops - Paris 2021
#1

Hi,

After Douglas Eck’s presentation of Magenta, Google’s TensorFlow offshoot and (market-oriented) research into generative kitsch, I’d like to pose a question on what I’d call the issue of euphony. I’m broadly and loosely calling euphony what makes Music sound good not only as it pertains to a particular canon or tradition, but also what makes Music sound good independently of canons and traditions. Here’s my question:

Is there any Machine Learning research being done to probe cross-canonical and cross-traditional corpora in order to discover Weltmusik (or Géomusique, as Deleuze would put it), i.e. the (hypothetical) generative topos/space of good-sounding Music independently of canons and traditions?

If, as I suspect, there is indeed (some approximation of) such a generative topos/space waiting to be discovered beneath the buoyant surface of Global Music History, I’d really love to explore it theoretically and creatively — to know and understand it of course, and ideally even to experiment with potentially new and unheard forms of good-sounding Music — with all the computational (Machine Learning) tools available to us.

Any hints or insights, as well as any fleeting/free-floating thoughts, on this issue would be much appreciated.

All the best,
António

1 Like
#2

Hi,

For anyone interested in piercing through the field of generative Music (and assisted Music creation), here are some recent high-grade references that can assist you (as they’re assisting me) in the journey.

Two papers by Jean-Pierre Briot (CNRS) and François Pachet (Spotify):

The deep dive in book form:

  • Deep Learning Techniques for Music Generation (Briot et al., Springer, 2019.)
  • Handbook of Artificial Intelligence for Music (Miranda et al., Springer, forthcoming 2021 — chapter/paper preprints available online.)

All the best,
António

1 Like
#3

Hi @antonioflorenca,

Thanks for starting this topic. DL is growing so fast and can solve issues in various music tasks. From MIR to generative music, via style imitation. Thanks for sharing these ressources (BTW do you have a link to Miranda preprints available online ?).
Another recent interesting paper is from Geoffroy Peeters, former Ircam MIR researcher. https://www.springerprofessional.de/the-deep-learning-revolution-in-mir-the-pros-and-cons-the-needs-/18945020

My 2 cents

#4

Hi Greg,

No problem, you invited me to repost my original question here, and I was perfectly happy to do it.

A few preprints of several chapters of HAI4M available online:

https://hal.archives-ouvertes.fr/hal-03081561/document

https://hal.archives-ouvertes.fr/hal-03046229/document
https://psyarxiv.com/a5yxf/

Thanks for Peeters’ paper, he did great work on timbre and low-level audio descriptors. Even if it’s not his field, it should be an interesting read.

All the best,
António

#5

Thanks for starting this topic.

#6

Hi @antonioflorenca,

Thanks for starting this topic. DL is growing so fast and can solve issues in various music tasks.

#7

Hi Antonio.

I’m just venturing out to try and add a bit to the Euphony story. I would respond here with an analogy to research conducted almost a decade ago (!) on the perception of emotions:
Universal and culture-specific factors in the recognition and performance of musical affect expressions.
https://psycnet.apa.org/record/2013-04185-001
Since Ekmann and his comparative studies with primates, it has been generally accepted in the field of emotion research that there are universal primary emotions and so-called cultural secondary emotions. Metaphorically speaking, this duality could be compared to the “innate-acquired” debate, since our universal primary emotions are considered to be close to our reptilian brain, while the others would be manifested by broader neurological activities, activating different brain areas. If we place ourselves in this field of analogy, we must then take into account epi-genetics and consider that from one generation to another, these boundaries evolve. For example, it can be assumed that if a piece of music is declared universally good at generation N, then it will certainly be good at generation N+4. We are therefore responsible for producing good music for future generations. One can ask the question of the success of a national anthem, or even that of Michael Jackson. Is it because his music is universal, or because he has been decreed King of Pop at some point?

#8

Hi Greg,

Culture and convention/conformity play a considerable role in both music creation and reception, that much we all know and agree on… But it still isn’t clear to what point any particular tradition/canon is a path of discovery of universals and not just a mere collage of socially enforced particulars.

The paper you linked seems to show that we should find a mix of both in every canon and tradition (as expected, I think), which means that there are indeed such cross-cultural universals.

The existence of such cross-cultural universals (or invariants) leads me to believe that Global Music History and Ethnomusicology (as a research field) clearly needs its own manifold hypothesis (as Peeters, in the paper you linked above in this thread, calls for in MIR.)

Thanks for the paper, it’s a gread read. There’s also a great review of this whole topic in Part I of the Oxford Handbook of Music Psychology (2015/16.)

All the best,
António

1 Like