an open spatial-audio playground

Omniphony

Throw sounds anywhere in the room, in real time. Watch them fly around a live 3D scene. Then rip out the panner and drop in your own — object-based spatial audio, built to be taken apart.

It puts sound in space. You get to mess with how.

Born one itch at a time — read the origin story →

Omniphony decodes object-based audio, works out where every sound should come from, and drives your real speakers with VBAP — or your headphones, through an HRTF or a measured room — live, with a 3D scene showing each object move and everything streamed over OSC. Nothing's a black box: formats load as runtime bridges, and the part that decides which speaker gets what is a backend you can rewrite in a single file.

  • No formats baked in — decoders load at runtime as bridges, one per codec family.
  • PipeWire on Linux, ASIO or WASAPI on Windows, CoreAudio on macOS, or raw float straight to a file or pipe.
  • Any number of speakers — 7.1.4, 9.1.6, or 128 if you've got the wall space.
  • Write your own panner in one file — it shows up in the UI on its own.
orender engineinputfile · pipe · mpvbridgesone per codec familybackendposition → gains← swap in yoursoutputsspeakers · phones · filestreamPCM + objectsfeeds

Fresh off the bench

What's new in v0.7.0

The highlights. The release notes have the rest — including what changed under your feet.

  • One Studio, with installers

    The native Studio is the only one now: .deb, AppImage, Windows setup and MSI, macOS .dmg. It opens on an Essentials view, and keeps the player’s engine up to date.

  • Everything in one installer

    Player, engine and Studio in one go on Linux and Windows — per user, no admin rights, with an uninstaller.

  • One bridge per codec family

    The engine loads every decoder bridge it finds and hands each stream to the one that recognises it. orender check-bridges says what it loaded.

  • Latency held by one number

    The adaptive resampler and its tuning wizard are gone. latency_target is the time from the renderer to your speakers, and one loop holds it.

  • Measured rooms on headphones

    Load a BRIR set and hear the studio it was recorded in, on its own speakers. Fixed channels land at their nominal angles on headphones.

  • Depth, and no speaker limit

    The volumetric backend adds an object’s depth inside the room to VBAP’s direction. The 24-speaker cap is gone — tested to 128.

Integration

Spatial playback in mpv

Playback guide

mpv-omniphony is the mpv media player with ad_orender — an opt-in spatial audio decoder that hands raw access units to the engine for object rendering instead of letting FFmpeg downmix.

  • Opt in with --ad=orender; plain streams keep their normal path.
  • Reads the same shared config as the CLI and Studio — and updating Studio updates its engine.
  • An AppImage for any Linux, ASIO built in on Windows, an .app on macOS.
  • Plays IAMF / Eclipsa Audio too, experimentally, on our own Matroska track: what works and what doesn't.

Come break things

Bring your own algorithm

Write a backend

This is the fun part. Copy the example backend, compute a per-speaker gain vector however you feel like — VBAP, DBAP, or something you made up at 2 a.m. — and the engine wires it into the render loop and the Studio UI for you. No central files to edit, no build system to fight.

Panic in the audio thread? A build-time smoke test catches a misbehaving backend before it ever reaches your ears. So go on: experiment, measure, break it, and tell us what you found.

A toy pairwise panner, live: drag the source and watch each speaker's gain follow. Your backend replaces exactly this decision — one function from position to gains.

Keep the tinkering going

Omniphony is free and open source under GPL-3.0 — a nights-and-weekends project, not a product. If it scratched an itch or saved you an afternoon, a sponsorship helps keep the experiments coming.

Curious how it got here? Read the origin story →