1 tinaudio
Sound matching asks how synthesizer controls can reproduce the audible shape of a reference. The current browser demo uses a conditional flow-matching model: reference audio plus either extracted or authored music-sketch controls generate 92 encoded coordinates, which decode to 89 Surge XT synthesizer parameters. Surge XT then renders the generated patch locally in the browser.
Figure 1: Current demo path: reference audio and extracted or authored controls feed flow inference, parameter decoding, and browser-native Surge XT rendering. The future FXP workflow enters through audio rendering, not direct source-parameter conditioning.
Figure 2: Planned control-track and trajectory visualizations. These placeholders are not presented as experimental results.
Virtual poster, Audio Developer Conference (ADC) 2026.
@misc{tinubu2026sketchcontrols,
title={Sketch Controls for Synth Presets: Editing a Surge XT Patch to Match a Target Sound},
author={Khaled Tinubu},
howpublished={Virtual poster, Audio Developer Conference (ADC)},
year={2026},
url={https://synth.tinaudio.com/}
}
The evaluator runs ONNX inference and Surge XT rendering in this page; audio stays in your browser. Upload reference audio, extract or author music-sketch controls, generate a patch, and listen to the real browser render. Loading is opt-in because the model and engine require about 200 MB of downloads. A short, low-step run checks integration rather than sound-matching quality.
Inference runs on the CPU via WebAssembly. The model generates synth parameters; Surge XT renders the audio. The current model does not preserve or stay near the parameters of an existing preset.
The model consumes stereo mel features from the reference audio and 386 music controls on 32 frames. Controls can be extracted from separate sketch audio or authored as a pitch activation contour with loudness, spectral-centroid, onset, and end values.
Classifier-free guidance can use content and sketch together or independently. A seeded RK4 sampler predicts the encoded parameter coordinates, which are decoded into a Surge XT patch and rendered by the browser engine.
The page reports integration diagnostics including normalized mel MAE, waveform RMSE, RMS levels, and peak amplitude. These measurements check a run's wiring and output; they are not a published perceptual evaluation or a preset-preservation result.
Curated listening samples and control-track plots are forthcoming. The live demo above provides the current supported audio path.
The open-source synth-setter repository contains the data pipeline, training and evaluation code, ONNX export, and browser evaluator used by this page.