Demo

Concept

RAVE — a variational autoencoder from IRCAM — compresses a sound corpus into a 16-dimensional latent space. NeoChucao lets a performer navigate that space from a MIDI keyboard in real time. Pitch maps to z0, velocity to z1, mod wheel to z2 for timbre morphing, and a CC knob injects randomness into the remaining dimensions.

Technical Detail

• —MIDI input maps to a 16-element latent vector fed to the RAVE decoder via nn.ar in SuperCollider. Each note spawns a Synth node decoding at audio rate with 50 ms lag smoothing on latent coordinates.
• —An 800 ms ADSR envelope hides the decoder's 200–500 ms warm-up artefact — a constraint of the convolutional architecture, not a musical choice.
• —Signal chain: Compander (~3:1 ratio) for dynamics, three-voice chorus to mask timbral wobble, FreeVerb2 for depth.

Iteration process

• —Python + Supriya: Python-driven architecture with mido, supriya, and a real-time 3D latent trajectory visualizer in vispy. Makes the geometry of the performance visible, but the added infrastructure pulled focus from the instrument itself.
• —live-coding-rave: Claude navigates the latent space through language via ClaudeCollider — an MCP server that generates and executes SuperCollider code live. Multiple RAVE models run in parallel on separate buses. Raises questions about authorship when the navigator is a language model rather than a hand on a keyboard.

Learnings

• —Moving from Python/Supriya to a single SuperCollider file shifted development time from infrastructure back to musical decisions.
• —Latent space is genuinely unpredictable — some regions are expressive, others are dead. That instability is a property of what the model learned, not a bug to fix.