VRCC

Also available in 简体中文, 日本語 and 한국어.

Speak into your microphone and your words appear in the VRChat chatbox as live captions, with translations into up to three languages underneath. Everything runs on your own Windows 10 or 11 PC. No cloud services, no API keys.

You say:   "Hello, how are you today?"
Chatbox:   Hello, how are you today?
           こんにちは 今日はどうですか?

VRCC is built to be plug and play. Run it and the first-run wizard sizes up your machine, picks the models, and sets the performance mode for you; every recommendation traces back to a measured benchmark run rather than a guess. The tuning knobs are still there in Settings, but you should not need them.

Get VRCC running

This takes about five minutes, plus a one-time model download. There is no Python to install and no API keys to find; everything runs locally on Windows 10 or 11.

1. Open the Releases page. The ready-to-run program is not on the repository page you land on first. It sits behind Releases. That link goes straight to the newest one. You will see a version number, the release notes, and at the bottom a list headed Assets. The files live in there.

   The last two entries under Assets, "Source code (zip)" and "Source code (tar.gz)", are VRCC's own code rather than the program. Skip both. The two you are choosing between are named VRCC-cuda-windows-x64 and VRCC-windows-x64, each followed by the version number.

2. Take the zip that suits your PC. VRCC-cuda-windows-x64 is the CUDA build, for an NVIDIA graphics card with 8 GB of VRAM or more on driver 570 or newer. Captions are worked out on the card and appear almost at once. VRCC-windows-x64 is the CPU build. It is a much smaller download and the one to take on an NVIDIA card under 8 GB, on AMD or Intel graphics, and on a PC with no graphics card at all. Captions are identical, just a moment slower, and the default models are sized to keep up on CPU.

   To see what you have, press Ctrl+Shift+Esc for Task Manager, open the Performance tab and click GPU. The card's name, its dedicated memory and the driver version are all on that page. If you would rather not look, take the CUDA zip. It is the bigger download, but it runs on the processor by itself when it finds no card it can use.

   GPU acceleration covers NVIDIA cards only for now, because there is no AMD hardware here to test on.

3. Unblock the zip, then unpack it. Right-click the file you downloaded and choose Properties, at the bottom of the menu. If there is an Unblock tickbox near the bottom of the General tab, tick it and click OK. Windows marks files that came from the internet, and clearing the mark here also saves you the warning in the next step.

   Then right-click the zip again and choose Extract All. Explorer shows a zip as though it were a folder, but starting the program from in there hands Windows one file out of the middle of it and it will not run. Pick somewhere like Documents and you get a VRCC folder with VRCC.exe inside. Keep that folder together and put it wherever you like. There is nothing to install.

4. Run VRCC.exe. Double-click it in the folder you just extracted. Windows may put up a blue box headed "Windows protected your PC". Click More info, the small link under the message, then the Run anyway button that appears. VRCC is not code signed, so Windows shows this for a program it has not seen much of. That box is worth clicking through for the file you took from the Releases page above, and it is not advice to click through it anywhere else. If you ticked Unblock a moment ago, it may not appear at all.

5. Let the first-run wizard download the models. The first time you launch VRCC, a wizard checks your machine and picks a speech-to-text model and a translation model for you (about 1 to 3 GB to fetch), then downloads them. You can switch the run device or the models before it starts. This only happens once. First run has the table of what it picks and why.

6. Enable OSC in VRChat. Open Action menu → Options → OSC → Enabled. VRCC sends captions to VRChat over OSC (default address 127.0.0.1:9000); without this, they won't reach the chatbox.

7. Choose your microphone and languages. In VRCC, pick your microphone, your source language (the one you speak), and up to three target languages to translate into.

8. Start talking. Your words appear in the VRChat chatbox as live captions, with the translations underneath. You can also type into the box at the bottom of the window to send text the same way.

Later, open Settings to change the interface language, swap models, or switch between Speed and Quality modes. Usage covers the rest of what VRCC does day to day, and Troubleshooting is where to start if captions never turn up. Building from source is for developers who want to contribute (see DEVELOPING.md).

What runs on your machine

Your voice never leaves your PC. There is no cloud service behind VRCC and no API key to paste in.

Speech recognition runs through faster-whisper, or NVIDIA's Parakeet TDT 0.6B v3 (run as an ONNX export via onnx-asr) for European languages, or SenseVoice-Small for Chinese, Japanese, Korean and English, which is available but not recommended. Machine translation runs through CTranslate2, with the NLLB, M2M100 and MADLAD models. The wizard picks from these on your behalf, and Picking a model has the measurements behind the choice.

First run

If the configured models aren't downloaded yet, a first-run wizard opens, picks models for your hardware and your spoken language (taken from your Windows display language), and downloads them:

Your machine	Speech-to-text	Translation	Download
NVIDIA GPU with 16 GB+ VRAM	large-v3-turbo	nllb-1.3B-int8	~3 GB
NVIDIA GPU with 11 GB to 16 GB VRAM	large-v3-turbo	nllb-600M-int8	~2.3 GB
No GPU, or an NVIDIA GPU under 11 GB, a language Parakeet covers	parakeet-tdt-0.6b-v3	nllb-600M-int8	~1.3 GB
No GPU, or an NVIDIA GPU under 11 GB, any other language	whisper small	nllb-600M-int8	~1.1 GB

VRChat itself wants a lot of the card's video memory. VRCC reserves 8 GB for it and sizes the voice model against whatever the card has left once the translation model is accounted for, which is why large-v3-turbo needs roughly 11 GB rather than 8: under that, it would leave VRChat too little room to run alongside it.

Japanese, Korean and Chinese use the same models as everything else. Parakeet does not cover them, so a machine with no GPU, or an NVIDIA GPU under 11 GB, takes whisper small rather than Parakeet. sense-voice-small is still available in the Models window and is much smaller (240 MB), but faster-whisper transcribed real VRChat speech more accurately in testing, so it is no longer recommended.

The wizard shows what it picked and lets you switch the run device before downloading.

Other options range from whisper tiny (~75 MB) up to large-v3 (~3 GB), plus NVIDIA's parakeet-tdt-0.6b-v3 (~690 MB, very accurate and fast), limited to English + 24 other European languages (no Japanese/Korean/Chinese). MT models range from m2m100-418M-int8 (~480 MB) up to madlad400-3b (~3.5 GB). NLLB leads the recommendations on measured caption quality; the M2M100 models are the permissive-license alternative (see Model licenses). Models can be added/removed later via the Models dialog. See Picking a model below for measured accuracy and speed.

Once VRCC is ready to run, the main window opens beside a panel of setup steps, with rows for turning captioning on, the voice model finishing its load, your voice reaching VRCC, VRChat turning up on the network, and a caption going out over OSC, plus an optional row for sound picked up from your speakers that never blocks the rest. The VRChat row only proves its OSC service advertised itself over mDNS, not that anything reached the game, so a network that blocks mDNS can leave the row unticked while VRChat runs fine, and the panel goes on opening at every launch for as long as that lasts. The caption row is evidence in the other direction. It proves a caption was sent, not that VRChat picked it up, because OSC never confirms that a message got through, and it drops out of what the panel requires whenever sending to VRChat is switched off. The panel stops opening on later launches once every row it still requires has passed. Closing it before then dismisses it for that session only, and it is back on the next start. Its own Don't show this again button dismisses it for good instead, whether or not every row has passed. Settings, on the Simple page, brings it back either way.

Usage

1. Start VRChat and enable OSC: Action menu → Options → OSC → Enabled.
2. Start VRCC. Pick your microphone, source language and up to three target languages.
3. Talk. Your mic is cleaned of steady background noise first (on by default), then utterances are segmented automatically (Silero VAD), transcribed, translated, and sent to the chatbox, throttled to stay inside VRChat's chatbox rate limit so continuous speech never triggers the in-game spam mute. Turn it down or off under Settings → Voice recognition → Reduce background noise.
4. Read other people: press Hear others in the main window to transcribe and translate the voices coming out of your speakers, so you can follow someone speaking a language you don't read. Choose which device it listens to, and which language it shows, under Settings → Simple → What other people say. Two things to know. It captures the whole output device rather than VRChat's voice channel, because Windows offers no per-app voice tap, so game and world audio are transcribed too. And it is shown only in the VRCC window: other people's words are never sent to the chatbox, since the room already heard them and relaying it would republish their speech under your name. Without a graphics card it shares one voice model with your own captions, so both get slower.
5. Typed messages: the text box at the bottom of the main window sends typed text through the same translate → chatbox path (useful when you'd rather not speak).
6. Mute sync: when enabled, muting yourself in VRChat makes VRCC stop listening entirely (configurable to ignore or invert). Speech made while muted is never captured, so unmuting mid-sentence captions only what you say after the unmute. This uses VRChat's OSCQuery discovery and works only when VRChat runs on the same PC (localhost); captioning itself works regardless.

Interface language

The interface follows your Windows display language by default and can speak 18 languages (English, 日本語, 한국어, 简体中文, 繁體中文, Español, Français, Deutsch, Italiano, Português (Brasil), Русский, Українська, Polski, Nederlands, Türkçe, Bahasa Indonesia, Tiếng Việt, ไทย). Pick a different one under Settings → Simple → Language; it applies as soon as the Settings window closes. This only affects VRCC's own interface; caption languages are chosen in the main window.

Performance modes

Settings → Simple → Mode switches between two presets:

• Speed (default): greedy decoding (beam 1), and it finalises a caption after 600 ms of silence. Captions appear fastest.
• Quality: beam 5, and it waits 800 ms. Noticeably better phrasing, about 200 ms slower per caption.

Mode changes captions only. Translation keeps its own search width under Settings → Translation → Advanced (fine-tuning), because the measurements behind these presets cover speech recognition and say nothing about translation.

Switching mode rewrites the pause timings on the Advanced page and Search width on the Voice recognition page. If you have set any of those by hand, VRCC names them and asks first; flipping between the two modes to compare never prompts, because those values came from a preset rather than from you.

Parakeet and SenseVoice always decode at full accuracy, so the Mode control is greyed out while either is the active voice model. The individual knobs (VAD timings, beam sizes, quality gates and so on) live in Settings → Advanced.

Picking a model

Every figure in this section comes from tools/bench_stt.py: 100 LibriSpeech test-clean utterances run through the same engine path the app uses, on one machine (Windows 11, Ryzen 9 9950X3D, RTX 5090, driver 610.62). WER is word error rate on English read speech; latency is the median time to transcribe one utterance. The accuracy numbers and the relative speed ratios carry over to other machines; the absolute latencies do not. On a slower CPU, expect every CPU time here to stretch by roughly the same factor. The full tables and methodology are in DEVELOPING.md, and benchmarks/RESULTS.md collects numbers from other hardware.

On an NVIDIA GPU, keep the default large-v3-turbo. It matched large-v3 at 1.7% WER while being about 3.5x faster, and it handles every language.

On CPU it depends on the language you speak:

• One of the 25 European languages: use parakeet-tdt-0.6b-v3. It reaches 2.3% at 0.13 s, beating the small default (3.7% at 0.74 s) on accuracy and latency, and it is not close. It also detects the spoken language on its own within that set.
• Japanese, Korean, Chinese, or anything else outside that set: stay on small. Every whisper model that beats it needs seconds per caption on a CPU.

Parakeet is faster on the CPU than on the GPU (0.13 s vs 0.21 s), because its int8 ONNX graph does not suit CUDA. So the CPU build is enough for it, and if you play VRChat on the same PC it leaves the whole GPU to the game. VRCC does this for you when the device is left on Auto.

The distil models lost to large-v3-turbo on GPU and to Parakeet and small on CPU in these runs, so there's little reason to pick them.

The first-run wizard picks for your hardware and your spoken language, which it takes from your Windows display language. On a CPU that means Parakeet when you speak a language it covers, and small when you do not. Set the spoken language to Auto and it stays with the whisper models: with no language known ahead of time, a European-only model cannot be trusted to cover it.

Where things are stored

What	Default location
Config	%LOCALAPPDATA%\VRCC\VRCC\config.json
Models	%LOCALAPPDATA%\VRCC\VRCC\models\
Logs	%LOCALAPPDATA%\VRCC\VRCC\logs\vrcc-<date>-<time>.log

Every run writes its own log file at full debug detail, and the five newest are kept. When reporting a problem, attach the newest file from that folder.

Run with --portable to keep config, models and logs in the application's own directory instead (handy on a USB stick or for isolated installs).

Updating

VRCC looks for a newer release when it starts, and puts up a notice if it finds one. The notice has a button that opens the release page in your browser. Nothing is downloaded or swapped for you, so updating is the same few steps as the first time: take the new zip, unblock it, extract it, and run VRCC.exe from the new folder.

Your settings and your downloaded models do not live in the application folder (see the table above), so they carry over by themselves. The models are not fetched a second time and the first-run wizard does not reappear. You can keep the old folder until you are happy, then delete it.

If you would rather not be told, open Settings and go to the Advanced / Power users page, then untick Tell me when a new version is available.

A portable install works the other way round. Under --portable the config, models and logs all sit in the application's own folder, so replacing that folder throws them away. Copy config.json and the models folder into the new one first, or unpack the new build over the top of the old.

Model licenses

The code in this repository is separate from the models it downloads; check that a model's license fits your use:

• NLLB models (default nllb-600M-int8, plus 1.3B and 3.3B): CC-BY-NC-4.0, non-commercial use only. These are the recommended models, so if you use VRCC commercially (paid or sponsored streaming, or at work) pick an M2M100 model in the Models window instead.
• M2M100 models (m2m100-418M-int8, plus 1.2B): MIT, so they carry no such restriction. In a blind comparison over 60 VRChat-style utterances into Japanese, Chinese and Korean they were less accurate than NLLB 600M, which is the trade for the permissive license.
• MADLAD-400 (madlad400-3b): Apache-2.0.
• Whisper models: MIT (OpenAI weights, SYSTRAN CT2 conversions).
• Parakeet (parakeet-tdt-0.6b-v3): CC-BY-4.0 (NVIDIA weights, istupakov ONNX export).
• GTCRN (the Reduce background noise denoiser, bundled not downloaded): MIT (Xiaobin Rong, https://github.com/Xiaobin-Rong/gtcrn).

Troubleshooting

• "No GPU detected" / everything runs on CPU: make sure you're running the CUDA build (the VRCC-cuda-windows-x64 zip, or a source install with pip install -e .[cuda]) and that your NVIDIA driver is ≥ 570. GPU acceleration is NVIDIA-only for now (no AMD hardware to test on). CPU-only operation is normal otherwise: captions are identical, just a moment slower; the default models are sized to keep up on CPU.
• GPU runs out of VRAM: the engines fall back to CPU (int8) automatically for the session (same captions, higher latency); pick a smaller model to stay on GPU.
• VRChat isn't showing captions: enable OSC in VRChat (Action menu → Options → OSC), and check the OSC address in VRCC's settings matches where VRChat listens (default 127.0.0.1:9000). If you use other OSC tools (e.g. a router like OSCRepeater), point VRCC at the router's port.
• Mute sync does nothing: it requires VRChat on the same machine (localhost mDNS/OSCQuery discovery), OSC enabled in-game, and an avatar that reports MuteSelf.
• A small screen appears when VRCC starts: it names what VRCC is loading and closes once VRCC has something ready to show. The first launch after an update or a reboot is slower, because Windows is reading the program's files from disk for the first time. If a launch dies, the newest log file shows how far it got, which helps when reporting a problem.
• First transcription is slow: model load and warm-up continue inside the window after the startup screen closes; captions flow at full speed afterwards.
• Reporting a bug: attach the newest file from the logs folder above. Each run writes one file at full debug detail and the five newest are kept, so the file with the latest timestamp is the run that went wrong.
• Starting VRCC again does nothing: only one copy runs at a time, and launching it again asks the running copy to show itself. If it doesn't come forward, Windows may be holding focus for whatever you clicked last, so the taskbar button is where to look.
• Setup steps don't reappear: closing the panel dismisses it for that session, and it opens again next start unless the rows it requires have passed. Its own Don't show this again button dismisses it for good instead. Either way, open Settings → Simple and use Bring back the setup steps to get it back.

Developing

Building from source, tests, benchmarks and packaging live in DEVELOPING.md.