Synthesis bench

The same score, rendered five ways and loudness-matched to −18 LUFS. Switch between them at the same playhead with 1–5.

Style Seed Length Lead GM fallback

Pick a style and press Render.

0:00 / 0:00

Your ratings

How this works and where the sounds come from

All variants play the identical note list from the engine (seeded). Choosing a lead swaps the melody to a monophonic instrument on every variant, so the neural lead can be compared fairly.
Every variant except "OPL3 FM (today)" goes through the same per-style mix bus: stem level matching, high-pass cleanup, kick-driven ducking, chorus, synced delay, generated reverb, tape/vinyl for lo-fi, glue compression.
Samples: Splendid Grand Piano (public domain), VCSL vibraphone/marimba/tenor sax (CC0), Smolken double bass (CC0), Greg Sullivan Wurlitzer EP200 (CC BY 3.0), MusyngKite (CC BY-SA 3.0) or FluidR3 (MIT) GM soundfonts, LinnDrum / TR-808 / Sonic Pi kits; served by danigb/samples and gleitz/midi-js-soundfonts.
Neural: Magenta DDSP checkpoints (Apache 2.0) on TensorFlow.js WebGL. Pitch and loudness curves are synthesized from the score, not analysed from audio.
Render time is wall-clock on this machine, single-threaded on the page; "× realtime" above 1 means it could keep up live.