Computational Creativity — Autonomous Musical Agents & Procedural Scoring

Computational creativity at ZIAA is practice-led and system-evaluated: an agent is not a metaphor but a persistent software entity with memory, taste, and failure modes. Early agents (2021–22) used Markov and constraint logic; later generations employ latent audio encoders that traverse learned timbre spaces under joystick or gestural control.

The division’s signature artifact is the real-time neural latent audio resynthesizer — a multi-vector instrument that morphs continuously between source spectra without MIDI note quantization. Performers navigate a 2-D latent map; the system proposes voice-leading that the performer can accept, contest or redirect. A sister lineage explores autonomous procedural scoring for the Svalbard and Mojave arrays, where long-duration field recordings become training signals for site-specific generators.

Evaluation privileges durational listening and peer critique over benchmark scores. Every generative prototype is documented with audio profiles, failure post-mortems when agents diverge, and monograph treatment of aesthetic commitments.

Agents are evaluated by durational deployment, not single-output benchmarks. A new generative prototype must survive a 90-minute unsupervised session in the Mojave listening room without collapse into repetition or silence, then a 60-minute duet with a human performer who is instructed to contest, not cooperate. Telemetry — voice-leading entropy, latent-space coverage, intervention frequency — is logged alongside performer interview. Only then is a monograph drafted; ESSAY-2025-05 (“Autonomous Audio Agents”) and ESSAY-2024-03 (“Spatial Psychoacoustics”) both include negative results where agents diverged.

Software lineage is openly versioned. The 130-entry System Audit ledger tracks every firmware and DSP kernel change by commit hash, author and target record — a citable provenance chain that mirrors academic software releases. Browser instruments are the public-facing runtime: SYNTHESIS//SIGNAL’s 48-voice granulator and the Acoustic Bench’s zero-allocation AudioWorklet kernels are the same code that ran in the field, not a simplified demo. This continuity is why a Scholar citation to a monograph can plausibly include a link to a running system.