Listening
Say "Modbot" and then what you want — clip that, show overlay, hide overlay — so you never have to reach a keyboard in a headset. Off unless you switch it on, and nothing it hears is recorded, kept or sent.
You are in a headset, something happens, and Save a clip is on a panel you have to find and point at. Listening is the answer to that: with it on, you say the companion's name and then what you want — "Modbot, clip that", "Modbot, show overlay", "Modbot, hide overlay" — and it tells you whether it worked.
This narrows something the companion used to promise
Until 19 September 2026 the companion said, in its own source and on this site, that it never recorded sound, by any route, and the build failed if any code that could open a microphone appeared anywhere in it. That ban is now narrowed by one file for Listening (and, since 5 November 2026, by a second one for Dictate), and this page is the honest replacement rather than a patch over it.
What has not changed: nothing your microphone hears is recorded, kept or sent — not to your group's server, not to Modbot Cloud, not anywhere. The microphone is never part of a clip, so the one way the companion can send a clip, as the answer to a request, cannot carry anything it heard.
Clips are not silent any more, but that is nothing to do with your microphone: a clip carries the sound VRChat is playing, and that page says what is and is not in one. Your microphone is only ever opened by this switch, and only ever to hear one phrase.
It is off until you switch it on
A fresh install has it off. An existing companion updated into a version that can listen has it off. While it is off, no microphone is opened and nothing is asked of Windows' audio system at all — there is no listener, no phrase matcher and no extra thread.
Settings → Listening → Listen for "Modbot" is the whole of the consent.
It listens for its name, and only then for what you want
Nothing is listened for but the companion's own name. Sound goes to a matcher that has been given two spellings of Modbot and can answer nothing else at all.
Only once one of those has been heard is the matcher that knows the commands handed any sound — and only for the next five seconds. "Clip that" said across a table, or "hide overlay" in the middle of a sentence, reaches nothing, because there is nothing there for it to reach.
Saying "Modbot" and nothing else does nothing. It starts those five seconds; if you say nothing else, they end and the companion goes back to listening for its name. Nobody's clip is saved and no panel moves because somebody said the word.
While it is waiting, the companion says so — on the Listening card and at the top of every page of its window — and stops saying it the moment it would stop acting on you.
Leave a beat after the name
The companion cannot hear a word that was said before it started listening for one, because it keeps no sound. So "Modbot" and then, after a beat, "show overlay" is more reliable than running the two together. One name gets one command; say the name again for the next one.
What happens to the sound
Sound arrives from Windows in fifths of a second, is checked against the name — and, in those five seconds, against the commands — and is thrown away. That is the whole of it.
- Nothing is written to disk. Not a recording, not a transcript, not a log of what you said. The files allowed to open a microphone cannot open a file for writing, and the build fails if they learn how.
- Nothing is sent. That file cannot reach the network either. The only thing listening ever sends is the one download of the phrase matcher itself, described below.
- Nothing is held. At most a few seconds of sound sit in memory waiting to be checked, and they are overwritten by the next few seconds.
It is not transcribing you
The thing doing the checking is a phrase matcher, not a speech recogniser. It is a 3.3-megabyte model that is handed a short list of phrases and answers one question: was one of those just said? It has no ability to produce a transcript of anything else, and the companion does not ship a recogniser that could.
That is most of why it was chosen. The same library the companion already uses for the voice also offers a full speech recogniser; a program that runs unattended on your PC has no business carrying one.
When the microphone is open
Only while VRChat is running, which is the same rule Clips uses. The companion knows VRChat is running because lines are arriving in VRChat's own log — it never looks for a running program. VRChat closing closes the microphone. Quitting the companion closes it.
So the microphone is not open whenever the companion is. It is open while you are inside VRChat, which is when the thing it exists for can happen.
While it is open, the companion says so at the top of every page of its own window, for as long as it is open.
It does not take your microphone away from VRChat
Windows is asked for the microphone in shared mode — the same way VRChat, Discord and every other voice chat program ask for it. Windows mixes them and hands each program the same sound. Modbot never asks for the exclusive mode that would lock other programs out, and the build fails if that call appears.
If some other program has already taken your microphone exclusively, Modbot cannot open it, the Listening card says so, and everything else in the companion carries on exactly as before.
Choosing which microphone
Settings → Listening → Microphone lists the microphones your PC has. Pick your headset microphone when you are in VR and your desk microphone when you are not; change it whenever you like and the open one closes and the new one opens straight away.
Windows default is the first entry and is what you get until you pick something else. It follows Windows: switch a headset on and Modbot moves with it.
A microphone you picked that is not plugged in falls back to the Windows default rather than going quiet, and the card says so. Your choice is kept, so the moment it is back it is used again.
This narrows something else the companion used to promise
Until this list existed, Modbot asked Windows to follow the default microphone and never asked which microphones your PC had — the same as never asking for a list of your windows or your programs. Choosing one means asking for that list, so the promise is narrowed rather than left standing next to code that contradicts it.
It is narrowed in the same shape as everything else here. One file in the companion may ask, and it is one of the files allowed to open a microphone (the other is Dictate's); the build fails if a third one learns to. The list is only asked for while Listening is switched on — with the switch off, no microphone is listed or opened. (The client always looks at which speakers and headphones the PC has, for its own sounds and spoken announcements; that is a different list and it stays on the PC.) And what comes back is a name and an id for each device. Nothing is opened and nothing is heard by asking.
What you can say
The name, which is all it listens for until it hears one of them. Two spellings, because the matcher hears sounds rather than spellings and "Modbot" is not a word it was trained on:
- Modbot
- Mod bot
Then, within five seconds, one of these:
| What you say | What it does |
|---|---|
| clip that / clip this | Saves a clip, the same as pressing Save a clip |
| show overlay | Puts the headset panel back in front of you |
| hide overlay | Takes the headset panel away |
So: "Modbot, clip that", "Modbot, show overlay", "Modbot, hide overlay". Say it normally. "Clip this" means what "clip that" means.
Which overlay show and hide act on
The panel in your headset — the one in front of your face, which is what you mean when you say it while wearing one.
It does not touch the notification pop-ups: those come and go by themselves and are how you are told a flagged person walked in, so hiding them would lose the thing you most need. It does not touch the window over VRChat either — that one takes your keyboard when it comes up, which is right when you pressed a key for it and wrong when you said a word inside a headset, and it has its own shortcut.
"Hide overlay" is a curtain, not a switch. It does not change anything in your settings file, the panel is not taken down and rebuilt, and "show overlay" brings it straight back — showing what is happening now, not what was happening when it went away. If you want it gone for good, untick Overlay on the Settings page.
If the overlay is switched off altogether, or your headset is not running, it says so rather than pretending it did something.
The phrases cannot be changed
The matcher is not given words — it is given the small pieces its own vocabulary is built from, and working those out from a sentence needs a second library the companion does not ship. Every phrase above was worked out once, using the model's own tokenizer, with every piece checked against the model's vocabulary before it was written into the client.
Saying it twice does not do it twice
One spoken sentence is one action, and there are two rules holding that:
- One name, one command. The five seconds end the moment a command is taken. Anything else has to start again with the name.
- Nothing fires twice within six seconds, whichever command it was, which covers the matcher offering the same words twice as the rest of your sentence arrives and covers you repeating yourself because you were not sure it heard. It also means "Modbot, show overlay" immediately followed by "Modbot, hide overlay" does the first only — leave a few seconds between them.
What it says back
If the voice is on, it says it out loud:
| What happened | What it says |
|---|---|
| A clip was saved | Clip saved. |
| Clips is switched off | Clips are switched off, so there was nothing to save. |
| VRChat is not running | VRChat is not running, so there was nothing to save. |
| VRChat's window has not been found yet | Modbot has not found VRChat's window yet, so there was nothing to save. |
| The clips folder cannot be written to | The clips folder cannot be used, so nothing was saved. |
| The recorder had stopped | Recording has stopped, so nothing was saved. |
| The panel was put up | Overlay shown. |
| The panel was taken away | Overlay hidden. |
| The overlay is switched off | The overlay is switched off, so there was nothing to show. |
| The headset is not running | The headset is not running, so there was nothing to show. |
If the voice is off, something that worked plays the notification sound, and something that did not plays nothing. One short sound cannot say which reason it was, and a sound that meant both "done" and "not done" would be worse than silence.
Either way it is written into the companion's Audit Log page — including every time it heard you and nothing could be done — and for a clip the Save a clip control on the overlay says Clip saved or Clip not saved for eight seconds, the same as when you press it yourself.
A clip's answer arrives a moment after you speak, because the recorder writes the file on its own thread. If nothing has happened after ten seconds it is called a failure, which is deliberately generous: saying "not saved" about a clip that did land is the worse of the two mistakes. Showing and hiding the panel happens there and then, so those answer immediately.
The one download
Turning Listening on fetches the phrase matcher once, from one pinned address:
https://github.com/k2-fsa/sherpa-onnx/releases/download/kws-models/sherpa-onnx-kws-zipformer-gigaspeech-3.3M-2024-01-01.tar.bz2
17,626,723 bytes, SHA-256 f170013b4716e41b62b9bfd809687c207cef798ef9bc6534d524e17af9b6561aAbout 17 MB, with nothing attached to the request: no token, no id, nothing about you or your pairings. GitHub sees your IP address, as it does for any download. A download whose size or hash does not match exactly is thrown away rather than used.
It lands in %APPDATA%\Modbot\phrases, and only the six files that are actually needed are
written: the three parts of the model, its vocabulary, name.txt, and commands.txt. The archive's
spare 8-bit copies, its sample recordings and its readme are not written at all. Those last two are
plain text and you can open them: they are exactly what your PC listens for.
The model is Apache-2.0, from the sherpa-onnx project. It is listed in the repository's
THIRD-PARTY-NOTICES.md beside the voice.
What it costs
Small, and deliberately so — your frame rate in VRChat matters more than how quickly a phrase is noticed. The matcher is 3.3 million numbers, about a thousandth of the size of the voice, and it runs on one thread below normal priority. Nothing here has been measured on a machine running VRChat.
On Linux
Listening only works on Windows, for the same reason Clips does: the first thing it was built for was saving a clip, and there are no clips anywhere else. The Linux client says so on the card rather than quietly doing nothing.
Turning it off
Untick it. The microphone closes at once rather than at the next restart. Or set it in the settings file:
"listenForPhrase": { "on": false }The microphone you picked is kept in the same object, as Windows' own id for it. Leave it out for the Windows default:
"listenForPhrase": { "on": true, "microphone": "{0.0.1.00000000}.{…}" }Deleting the phrases folder under %APPDATA%\Modbot removes the downloaded matcher as well; it
is fetched again the next time you turn listening on.
Clips
Keep the last few minutes of VRChat and its sound while it is running, and save them as a video file — from the settings page or from inside your headset — or let the group's Modbot ask for a clip of a ban, kick or warning you made. Off unless you switch it on.
Dictate
Speak a question to AI Chat from the headset panel or the desktop overlay. Your words are written out on your PC, sent to AI Chat, and the answer shows in a Chat panel. Your voice is never saved or sent.