Unsupervised Wrap-Discontinuity Repair in Wavetable Synthesis via Hybrid GA–PPO Meta-Search

Cursor, Scispace, NotebookLM

PAPER · v1.7 · 2026-08-04 · ai

Formal Sciences Computer Science Artificial intelligence and machine learning

Abstract

In single-period wavetable synthesis, a sample discontinuity at the wrap boundary injects a broadband transient that repeats at the fundamental frequency, producing an audible periodic click. We address this cycle-local seam artifact with DenoiseOpt, a residualscored hybrid meta-search that selects and fits bake-cell operators applied at the period boundary. The bake cell composes classical DualCosine crossfades, FIR blends, and optional lightweight residual networks, possibly routed via Mixture-of-Experts (MoE) soft gating, without requiring paired studio-clean recordings. Candidates are ranked by the prolonged residual score R ∈ [0,1] (higher is better): the clamped tiled-RMS agreement between the baked multi-period output and a procedural ideal sibling that suppresses the open-wrap cliff. The outer search loop interleaves genetic algorithm (GA) tournament selection, PPO-based architecture proposals, and optional population-based training (PBT). On a frozen canonical sine+cliff holdout (seed 20260719), the compact metafavorite achieves R ≈ 0.991 versus DualCosine R ≈ 0.825 (∆R = +0.166). In an overnight GPU campaign (RTX 3090, 5,000 clean iterations), the champion reaches R ≈ 0.9909 versus DualCosine ≈ 0.8166 (∆R ≈ +0.174). Across twenty multi-family generative draws, mean ∆R = +0.162 over DualCosine (Wilcoxon p ≈ 8.9×10−5).

Keywords

Signal Processing Meta Learning Reinforcement Learning Edge AI Deep Learning Artifact Repair Denoising wavetable synthesis wrap discontinuity unsupervised repair seam restoration noise2noise neural architecture reinforcement learning genetic algorithms proximal policy optimization hybrid meta-search population-based training audio signal processing

Download PDF