Towards an Ecosystem of Instruments of Tunable Machine Learning

This paper introduces an embedded machine learning framework that integrates machine learning model training, tuning and control within live musical performance. Responding to critiques of disembodied and disempowering ML tools, we propose an ecosystem of Musically Embodied Machine Learning (MEML): low‑resource, self‑contained devices providing real-time machine learning on control data. We detail two open‑source board designs, a firmware stack featuring a C++ library for on‑device training of small networks (with focus on reinforcement learning) and an Arduino library that unifies real‑time audio, sensor I/O and ML tasks. Two reference instruments—a joystick‑driven FM synthesiser using interactive ML and a reinforcement‑learning variant—illustrate rapid prototyping and pedagogical value. The framework is being iteratively refined through university workshops and ongoing collaborations with professional musicians, who shape its future direction.