Everyone opens the same link and lands in the same room — no blobs. Share the URL from your address bar once you are in; it carries the room and passphrase.

not connected

Hollow dots are people you can see but are not subscribed to — no audio is being sent for them at all. That is the whole cost model: egress follows who is audible and talking, not who is in the room. Drop K or R and watch them go hollow.

↑↓ walk · ←→ turn your head · F face whoever is selected. Dragging still works for jumping somewhere.

Turning matters more than it sounds: you focus on a voice partly by putting it on-axis and swinging your head shadow onto the competing ones. A fixed heading removes half the mechanism.

Hold W to whisper to the selected source (click a dot). Its voice moves close and off-axis; everything else ducks. Release to return.

In a real room you cannot hear 20 m away mostly because ambient noise masks it. Headphones have no such floor — turn the volume up and a very quiet voice is still perfectly clear. Without this, distance is only quieter, never inaudible.

Low-pass is an intentional perceptual exaggeration, not physics — air absorption is negligible at office distances. Tune it by ear; that is the point of this rig.

The PRD gate: freeze the parameters, then score on material you did not tune with. Copy this before you start scoring — it is the candidate pipeline the Phase 0 exit gate asks for.

Direct browser-to-browser — no server in the media path, so the latency is real. Both wear headphones.

You: create invite → copy → send it over Slack.
Them: paste it → accept → copy the reply → send it back.
You: paste their reply → accept.

Just paste whatever you were sent and press accept — it works out which side you are.

Turn this off only once both of you are definitely on headphones — it smears the spatial cues, so the purest test wants it off, but with it off a speaker user will produce real feedback. A limiter and Esc panic mute sit in front of your ears either way.

not connected

Three real conversations, six voices, turn-taking with silences — each track holds only that speaker's turns. Pairs are placed adjacent so they interleave like a room.

Drop in real recordings for anything you intend to score — babble is a masker, not speech, and intelligibility needs words.

Listener is the ◎ at centre, facing up. Drag it or any source. Headphones required — HRTF over speakers is meaningless, and an open mic will feed back.