A MIDI loop goes in. A responsive musical partner grows around it.
Neuraloop is a work in progress: an instrument concept for turning a played MIDI loop into complementary musical parts, while keeping the performer in control.
Work in progress
The interaction and split hardware architecture are being designed. Model integration and the playable hardware prototype are still ahead; real-time co-performance is the goal, not a completed capability.
Inspiration and learning
I’m developing Neuraloop to find musical inspiration and learn how to train a model on musical patterns rather than language.
Design & implementation
The intended interaction starts with a played MIDI loop. The model proposes complementary parts that the performer can control and transform while playing.
The system is designed around a split architecture. An embedded ESP32-S3 orchestrator manages MIDI routing, controls, and system state in real time, while a separate compute node handles symbolic note generation. This keeps timing-critical behavior close to the hardware and pushes heavier inference away from the performance layer.
On the model side, Neuraloop is moving toward a role-aware generation approach instead of a generic LLM setup. The training direction combines groove-oriented datasets for rhythmic parts, broader melodic datasets for harmonic material, and multitrack datasets to learn arrangement structure and inter-part relationships.
Design goals
- Real-time MIDI co-composition centered on performer control
- Split architecture between embedded orchestration and remote inference
- Role-aware multitrack generation for drums, bass, synth, and pad
- Focus on low latency, interpretability, and live playability
- Training strategy built from complementary symbolic music datasets
