Reflector: open-source AI adapts sample retrieval to your arrangement

Reflector, developed by Austin Rockman, is a new interactive audio workstation designed for sample-based composition. Unlike traditional sample retrieval tools that query from a fixed reference, Reflector dynamically adapts to the evolving harmonic context of a composer's arrangement. The system is built around a fixed interval-class oracle—a hand-designed table of weights that scores pitch-class compatibility between sources. An encoder trained purely on synthetic audio approximates this oracle in a 128-dimensional embedding space, where dot products serve as compatibility scores at interactive speeds.

As the composer layers material on a multi-track timeline, Reflector performs sweep-line analysis to identify co-sounding regions, computes oracle-weighted centroids, and retrieves samples against the composite harmonic identity of the session in real time. Session centroids can be projected into a navigable 3D space to reveal structural harmonic relationships across a composer's body of work. The entire pipeline runs locally with no copyrighted training data, and both the application and training code are open source. Key findings show the learned embedding preserves the oracle's pairwise judgments while covering the full library, overcoming limitations of direct scoring methods.