Writing
Notes on the STRATA II engine
11 August 2026
Three RAVE models in continuous feedback, corrected by a listener module that measures the system's drift against its own untouched source.
STRATA II is driven by three instances of RAVE — Realtime Audio Variational
autoEncoder, Caillon & Esling, IRCAM ACIDS — running in Max 9 through the nn~
external. Each instance is wired as a feedback loop: its own output is returned to
its input, so degradation accumulates continuously as the piece plays rather than
in discrete generations.
Left alone, a system like this has two failure modes. It either runs away into noise, or it settles into a fixed point and stops moving. A listener module prevents both. It measures the spectral drift of the system’s current output against the untouched original and uses that distance to set the depth of erosion, so the piece is held in the region where the source is still recognisable and still being lost.
The three voices are encoded into a third-order ambisonic field and decoded to twelve loudspeakers — eight at ear level, four overhead. Elevation is tied to erosion depth, so as a copy takes over from the original it physically rises above it.
None of the three models was trained on the source material. One was trained on magnetic field recordings, one on a hand drum, one on roughly ten thousand stock sound-library loops. This is a compositional decision, not a limitation: each model can only rebuild Istanbul in terms it already possesses, and the gap between what it reconstructs and what was there is the subject of the piece.