Voice translation
for VRChat

No separate translation app is required for the automatic setup. SayWhat? includes its capture host and translator runtime, then prepares speech support and downloads models after you approve setup. Existing compatible installations can be reused from Advanced.

Yes. Languages offers a primary language and up to two additional languages, off by default. Translations appear on separate lines in one chat bubble. They reuse one loaded translator, but take additional inference time and share the chatbox text limit. The selected model must support every language; Liquid's EN–JP model cannot translate to Chinese or Korean.

Optional early messages are cumulative translations of the sentence so far. They are submitted chat bubbles, not keyboard drafts. The recognizer keeps the entire utterance as context and the completed sentence supplies the definitive translation. Updates are coalesced to roughly 1.5 seconds or slower; chatbox text is limited to 144 characters and 9 displayed lines. Long messages show a readable tail, not the whole paragraph at once.

Not as independent live audio streams. Optional voice recognition can help attach names to clear speech, but recognizing a person is different from separating overlapping voices. Overlap, music and recordings can still confuse recognition and translation.

Delay includes the pause after speech, recognition finishing, translation and the chatbox send cadence. VRChat and local models also share GPU resources. Try the smaller translator, one shared model, or just one translation direction. Advanced has timing and troubleshooting information; a larger model alone does not solve every delay.

SayWhat? asks whether to quit and release its model memory or keep running in the system tray. You can remember your choice and change it later in Settings. Closing the subtitle window stops subtitles independently.

The app checks for updates on launch unless you turn this off. Update now downloads and verifies the installer, then opens its update wizard; final confirmation and Windows permission remain visible. This path keeps your settings and models. The standalone installer also offers update, repair and uninstall with optional data choices. All-users operations keep each user’s private data.

Open VRChat’s Action Menu and select OSC → Enabled. SayWhat?'s sidebar checks local OSC availability and offers guidance if it cannot verify a connection. This is not a chat-delivery acknowledgment: chatbox visibility and user safety settings can still affect what other people see. See VRChat’s official OSC guide for current instructions.

Open Advanced → Diagnostics in the app. Note which direction stopped, your selected models and the timing shown, then report the problem on GitHub. Review logs for private conversation text before sharing them.