Origin story

One itch at a time

Omniphony is a real-time spatial player — format-agnostic, and built to be taken apart. It didn't start as a product; it started as a chain of itches, each one scratched just enough to reveal the next. Here's the short version.

  1. 01

    The positions were right there

    An open-source decoder for a well-known object-based cinema format had just appeared — and with it, something irresistible: the position of every sound object, live, as the stream plays. We wanted to hear what VBAP would do with that on real speakers.

  2. 02

    First try: an external panner

    The first experiment wired those positions into SPARTA's panner over OSC. It worked, and it sounded promising — but decoding and rendering lived in separate processes, and the OSC hop put latency between where a sound was and where you heard it.

  3. 03

    Rendering moves in-engine

    So, with some AI help, we brought the rendering inside the engine, applied just in time: positions and audio stay in one process, sample-aligned. No hop, no drift — the panner reacts the moment the object moves.

  4. 04

    Beyond the sphere

    VBAP renders on a virtual sphere, but objects like to live inside the room. We added alternative renderer backends to experiment with that — and then made backends a public, swappable part of the design, so anyone can bring their own algorithm.

  5. 05

    The pipe era

    The first player integration read audio from a named pipe and played out through PipeWire or ASIO, fed by mpv passthrough over IEC 61937. That meant giving up the player's resampling and owning audio sync ourselves. The mode still exists — it may yet earn its keep for network rendering.

  6. 06

    A decoder inside mpv

    Later we added an audio decoder to mpv itself — ad_orender in mpv-omniphony — so spatial rendering rides mpv's normal audio output, resampling, sync and all. That's the playback path most people use today.

  7. 07 · v0.5

    Tuning a room by ear

    Hearing is believing, so we built the test bench in: pink noise through each speaker's own crossover, and a test object you can park anywhere in the room to hear where it lands. Sounds panned outside the speakers stopped fading away and got virtual poles instead, the headphone stage learned to wear a whole speaker rig as a virtual room, and the mix lost every allocation and lock — with a CI gate to keep it that way.

  8. 08 · v0.6

    Speakers where the format puts them

    Not every format agrees on where a "height" speaker lives. Bridges now declare their formats' speaker positions, and each family of sources gets its own placement — on the sphere, in the room, or wherever you drag it. The binaural stage was rebuilt from the measurements up, and a native Studio showed up: same scene and knobs, no web view, idle at zero CPU.

  9. 09 · v0.7

    One Studio, one installer

    The web Studio retired, and the native one got proper installers. A single download now lays down player, engine and Studio together. The engine loads one bridge per codec family and hands each stream to the one that knows it, and the whole adaptive-resampling machinery gave way to one number: the latency you want, held by one loop.

  10. 10 · v0.7

    Borrowing other people's rooms

    Load a measured room — impulse responses recorded on a dummy head in a real studio — and hear your films in it, on its own speakers. Back in your own room, the volumetric backend finally gives objects depth inside it, as step 04 wanted, and the speaker count is no longer capped: we test up to 128.

There's a lot more in here than this story covers — that's what the docs are for. And if the next itch is yours, the source is on GitHub.

Grab a buildWrite a backend