AI that plays the modular synth.
DeepSynth turns text, audio, images, and live room context into a playable synth state — an interpretable graph of modules, cables, and parameters, verified before it reaches the artist.
How it works
Context in. A playable synth state out.
DeepSynth is not a song generator. It operates the instrument: it reads the moment, proposes the next synth state, and lets the artist audition, accept, or steer.
Listen
Prompt, audio, image, video, and the current PatchBee state become one compact performance context.
Compose
DeepSynth plans roles and decodes them into a modular synth graph — modules, cables, parameters, macros.
Verify
Every generated state is rendered and checked before it reaches the artist. Only states that sound play.
Architecture
Three tiers, one rule.
Graph decoding goes to small, fast models. Meaning goes to the large model. A verifier sits in front of everything — a state that doesn't sound never reaches the stage.
Live
under 1 second
Tiny role models decode synth graphs in real time — basslines, drums, leads — with deterministic assembly and repair.
Interactive
5–30 seconds
A large composer model plans the set: role selection, arrangement, and interpretation of artist feedback.
Offline
between shows
Artist feedback becomes training data. Models retrain against a verifier fleet so the next set starts better.
The output is never a black-box waveform. It is an interpretable modular graph the artist can open, re-patch, and play — with provenance for every module and cable the model touched.
Integrations
One engine. Any synth.
DeepSynth generates synth states, not audio files — so the same engine can drive any instrument it learns to speak. PatchBee is the supported integration today, with VST and DAW targets on the roadmap.
PatchBee
live · verifiedAI-native modular synth built for live patching and native DSP. The flagship integration: every generated state is rendered and verified for sound before it reaches the artist, and feedback flows back into training.
- Host
- VST / AU / native
- Input
- prompt + context
- Output
- playable state
VST instruments
coming soonGenerated states exported as native preset formats.
DAW workflows
coming soonGenerated states placed directly into project sessions.
API
Synth states, one call away.
Send a performance command; get back a routed plan and a playable synth state with quota metadata on every response. The same endpoint powers the live demo on this page.
curl -X POST https://api.lsm.sh/v1/generate \ -H "Authorization: Bearer dsk_live_…" \ -H "Content-Type: application/json" \ -d '{"command": "dark warehouse acid set"}'
{
"route": "conductor",
"liveSet": {
"patch_id": "patch_44ba3e5e48",
"roles": ["kick", "acid_bass", "closed_hat", …],
"patch": { modules, cables, params },
"sound_ok": true
},
"quota": { "freeRemaining": 4, "freeLimit": 5 }
}The artist
Can machines feel the room?
Matthew de Chang is an engineer and artist in San Francisco. After fifteen years of building distributed systems, he now builds instruments: AI that doesn't just generate music, but performs it — reading a room, patching a synthesizer, responding in real time.
code · voltage · sound — san francisco, ca
PatchBee
An AI-native modular synth. Computer vision reads the crowd; the patch responds.
DeepSynth
An AI that operates synthesizers — choosing modules, connecting cables, shaping timbre from voltage.
LSM
A Large Synth Model shaped by his own writing, research, and live performance practice.
Live demo
Text command in. Sound out.
Describe the moment in text; DeepSynth returns a routed plan and synth state, and the mini rack below plays it in your browser. Audio, image, and video context are coming soon.
/api/multimodal-preset
A synth state is the next musical decision.
Upload media or use the demo room. DeepSynth summarizes the performance context, chooses roles, maps macros, and requests a playable state from the routed runtime.
Mini rack
A tiny browser modular synth. Generated states are applied to it — press play to hear them.
Get early access to DeepSynth.
An artist-specific model that learns how a performer wants to move a room. Join the list for early access and release notes — open to artists and researchers.
Or write directly: info@deepsynth.ai