Surge music-sketch evaluation

Local browser inference and rendering. Audio stays in this page. Experimental, unchanged Surge 1.4 port; the model was trained with Surge 1.3.4. This is not a parity or model-quality claim.

Inputs become four-second stereo clips at 44.1 kHz, clipped to [-1, 1] like the native sketch loader.

Inputs Content / target audio Music sketch Sketch audio

Sampling WebGPU neural inference

Experimental: conditioning and velocity networks only. Mel, PESTO and Surge stay on WASM. Keep inputs, seed and settings unchanged for a previous-run parameter delta. Neural timing excludes model loading, feature extraction and audio rendering; the first run may compile shaders. Unsupported WebGPU reports an error, not silent fallback.

Conditioning Content CFG Sketch CFG RK4 steps Seed

Loading model and music frontend…