Skip to main content

VRCT v3.6.0

Β· 6 min read

VRCT v3.6.0 adds two new ways to use VRCT: Chat Transcribe, which reads VRChat chat bubbles on screen, and the VR UI (Beta), which lets you operate VRCT from inside VR. It also adds VAD-based voice detection, a redesigned updater, a searchable settings screen, and several security and stability fixes. This post covers the changes since v3.5.1.

New features​

Chat Transcribe (OCR)​

  • Reads the chat bubbles shown in VRChat and translates them through the usual translation pipeline.
  • Results appear in the message log and the SteamVR overlay. They are never sent to the VRChat chatbox.
  • While SteamVR is running, the VRChat view is read from SteamVR. Otherwise, the VRChat window is read.
  • Turn it on and off with the Chat Transcribe switch in the main window.
  • The new Chat Transcribe tab lets you choose:
    • Source language: Auto (Japanese, English, Chinese, and other Latin-script languages), or Korean, Russian, Ukrainian, Thai, Arabic, or Hindi. Changing the language while it runs switches without a restart. Arabic accuracy is currently low.
    • Window name (default: VRChat), for clients with a different window title.
    • Capture interval (default: 750 ms), Minimum confidence (default: 0.85), and Minimum text length (default: 2 characters).
  • If it cannot start, for example because the model is not ready, VRCT shows a notification with the reason.

VR UI (Beta)​

  • A launcher on your left wrist (you can switch hands) turns Translation, Microphone, Listen, and Chat Transcribe on and off.
  • Floating windows show the Log, Language, and VR settings.
    • Move, push, pull, and resize each window with your controller.
    • Anchor a window to the world, your left or right hand, or your head, and lock it in place.
    • If a window goes out of view, use Recall.
  • Log text size ranges from 14 to 28, and the opacity steps through 100%, 80%, 60%, and 40%.
  • Use the Language window to switch presets, swap your language with the first recipient language, and add or remove up to three recipient languages.
  • While the VR settings window is open, translation and transcription pause, and they resume when you close it.
  • Turn it on with Show the VR UI in the VR tab. If the launcher is lost, use Reset the launcher position and size.

Voice detection (VAD)​

  • The mic and speaker sensitivity modes now include VAD (Auto). An AI voice activity detector (Silero VAD) finds speech and sends each utterance to transcription when you finish speaking. This helps when background music or game sound is playing.
  • Each of the mic and speaker has its own setting. Choose VAD (Auto), Auto, or Manual.

Updater​

  • The update screen is redesigned. In one place you can choose:
    • The release channel (Stable or Beta)
    • The build (CPU, or GPU (CUDA))
    • The version to install, including earlier releases from v3.6.0 onward
  • Before you install, a table compares the current setup with the setup after installation. Warnings appear for downgrades, channel switches, and the GPU build, which needs about 5 GB of additional download.
  • The downloaded installer is checked against its published SHA-256 checksum. If the checksum does not match, the update is cancelled.
  • When you start the installer from VRCT, the installer language follows VRCT's display language.

Settings​

  • Search: Type in the search bar at the top of the Config Window to find a setting.
  • Translation tab: grouped into AI Translation (Local), External Translation API, and Local LLM & Compatible Server.
  • Device tab: the mic and speaker settings are split into their own sections.
  • Only the options that apply to the selected engine are shown. For example, the Whisper-only advanced settings appear only when Whisper is selected.
  • The "Auth Key" label is now "API Key" throughout.
  • You can press Enter to save a text entry. Number fields accept the arrow keys for fine adjustment, and the mouse wheel no longer changes their values by accident.
  • The model lists show a Default label on the default model.
  • If a hotkey is already used, VRCT shows an error instead of silently failing.

Improvements​

  • Startup checks the translation and transcription engines in parallel, so VRCT starts faster.
  • Translation into several target languages runs in parallel.
  • If you keep speaking without a pause, VRCT still splits the audio about every 15 seconds, so transcription does not fall behind.
  • When VRCT starts before VRChat, the mic mute state is synced each time you load an avatar in VRChat.
  • The mic's automatic selection follows the Windows default device. VRCT no longer switches to a virtual audio device on its own.
  • The VRC Mic Mute Sync and Set To Stay On Top switches in the top bar toggle directly, without a pop-up.
  • Notifications appear when you open the conversation log or settings folder. Toast error messages can be selected and copied.
  • Dropdown menus are no longer clipped by their parent element, and they close on an outside click or the Esc key.
  • The About VRCT documentation link opens in the language you are using.
  • The GPU build no longer depends on PyTorch.
  • The built-in web translation engines (Google, Bing, and Papago) were reimplemented. VRCT can start offline and translate once the network returns.
  • AI model downloads retry and time out when the connection drops, so a temporary network error no longer shows "model not found".

Security​

  • Downloaded updates are verified with SHA-256, and the update is aborted on a mismatch.
  • The WebSocket server requires an access token. Use WebSocket URL (click to copy) in the Advanced Settings tab to copy the URL with the token. The OBS Browser Source connects with the token automatically.
  • The WebSocket host can no longer be set to a wildcard address such as 0.0.0.0 or ::.
  • API keys are no longer written in plain text to logs or the standard output.

Bug fixes​

  • The OBS Browser Source no longer goes blank after an abnormal WebSocket disconnect.
  • Recording no longer stops and fails to resume after audio devices are unplugged or auto-selection is toggled.
  • A microphone set to WASAPI stereo input (for example, MOTU M Series "In 1-2") no longer shows a working level meter with no transcription.
  • Startup no longer hangs indefinitely, and VRCT can be closed while it is hanging.
  • Startup no longer waits for the GitHub update check.
  • Heavy log output no longer freezes the whole app for tens of seconds.
  • A clipboard paste that cannot focus the target window now shows a notification.
  • Invalid color values are no longer saved.
  • Clearing the OpenAI-compatible URL no longer wipes the other settings.
  • The Config Window no longer crashes when the selected processing device is missing from the list.
  • Slider tick marks and labels are shown again in the settings.
  • A rare parse error in the communication with the UI is fixed.

Removed​

  • The plugin feature has been discontinued. The Plugins tab and the plugin window are gone.
  • The poster section and the localization credits in About VRCT have been removed. Localization contributors are listed in the Special Thanks section.

The v3.5.1 documentation is available from the version dropdown in the top navigation bar.