Live coding RAVE neural audio synthesis models using TidalCycles and SuperDirt.

https://github.com/elgiano/nn.ar

Note: if you have a previous version of rave-supercollider installed in your SuperCollider Extensions folder, delete it, otherwise you will experience Symbol not found errors for libtorch.

• Add the RAVE OSC target (rave.hs) to your TidalCycles Boot.hs.
• Run rave.scd after SuperDirt has started.
• Test the patterns in rave.tidal.

You can find various CC-licensed models here: https://huggingface.co/Intelligent-Instruments-Lab/rave-models

You can also train your own models, see https://github.com/caillonantoine/RAVE.

For faster training via transfer learning, see https://github.com/victor-shepardson/RAVE.

This is a simple way of using RAVE in SuperDirt and is quite limited.

• RAVE models are mono.
• One RAVE model is assigned to one orbit.
• rave.scd replaces SuperDirt output buses of Orbits 0, 1 and 2 with it's own bus.
• No DirtPan currently interated into rave.scd.

• Most of rave.hs is only necessary because of rave.scd. If the latter can be improved, then only the custom parameters would be needed in Tidal.
• Not possible currently to pattern the RAVE model itself due to the SuperDirt limitations described above.
• RAVE latents not currently implemented on Tidal side.

• Improved integration with SuperDirt
• RAVE Prior variations
• Mixing and match encoders and decoders
• Patterning model names from Tidal (via Ndef?)

https://scsynth.org/t/experiences-with-rave-ugen/6398/32

https://club.tidalcycles.org/t/tidal-x-rave-realtime-audio-variational-autoencoder/4973