Local Autoregressive Symbolic Music Generation

A companion to the retinal segmentation work, applying the same local autoregressive idea to a domain where structure is explicit: polyphonic chorale generation. The framework models latent transitions locally, adds residual modelling, and introduces a gated drift mechanism to stop the generation wandering away from the harmonic context.

Where the imaging version memorised, this one did not — the model showed low memorisation, high diversity, and stable generation across random seeds. Self-similarity analysis and ablations against no-drift and no-gate variants isolate what the gating mechanism is actually contributing.