A Dataset and Analysis Framework for Large‑Scale Synthesizer Preset Generation
Abstract
This paper introduces the Sylenth1‑10KPresets Dataset, a corpus of 10,292 synthesizer presets paired with perceptual timbre descriptors. The dataset combines 292 factory presets with 10,000 parametrically generated variants, created through a factory‑conditioned randomization procedure that preserves audibility and musical plausibility. Each preset includes complete Sylenth1 parameter values and seven AudioCommons timbral descriptor values (brightness, warmth, hardness, depth, sharpness, roughness, boominess) computed from rendered audio. Statistical analyses characterize parameter distributions, descriptor trends, and cross‑model relationships. Principal component analysis reveals interpretable latent dimensions linking low‑level controls to perceptual timbre qualities. We identify distinct distributional patterns between expert‑designed factory presets and algorithmically generated sounds, revealing that factory presets occupy compact and structured regions of the parameter space while random presets explore broader parameter ranges. Three released baselines on a single set of recovery‑based, leakage‑free splits provide reference numbers for parameter‑to‑descriptor regression and descriptor‑to‑parameter inversion. All code, analysis scripts, model checkpoints, and metadata are released for full reproducibility, with the aim of supporting research in supervised timbre modeling, interpretable neural synthesis, and semantic sound‑design interfaces.
© 2026 Joseph M. Cameron, Alan F. Blackwell, published by Ubiquity Press
This work is licensed under the Creative Commons Attribution 4.0 License.