Caption-to-audio, from scratch

A sample is a file.
This is a function.

Type a sound. A 5M-parameter planner turns your caption into a typed JSON score, and plain NumPy renders it to 48 kHz stereo, exact to the sample. Both run here, in your browser.

engine idle

Exact captions, read by the planner

Or plain words, read by the planner itself

The planner (5.1 MB, int8) downloads on your first Generate. Sounds render with the NumPy DSP path; the 9M detail model that adds texture on top stays in Python, so the cues further down can sound a touch richer.

Shape it by hand

Every sound is a score you can edit.

Pick a score, then drag a marker or a slider. Your browser re-renders it with the same engine, no model in the loop.

Spec engine idle

            
          

Listen

Cues for real work.

Forty-one cues for apps, edits, brands, games, nature and everyday life. Each one is a single JSON score, rendered without samples or recordings. The physically modelled sounds are pure DSP; the detail model adds texture to the rest.

For agents

Built for agents.

Add ORB as an MCP server and your agent can score a scene, check the audit, and patch one field instead of regenerating.

{
  "mcpServers": {
    "orb": { "command": "orb-mcp" }
  }
}
  • 5.1Mplanner parameters, 20.4 MB. Runs on this page.
  • 9.1Moptional detail pass, 36.3 MB, CC BY-NC 4.0. Python only.
  • 48 kHzstereo, every event placed to the sample
  • NumPyis all the DSP renderer needs