Offline voice input
Use SenseVoice to recognize Chinese, English, Japanese, Korean and Cantonese locally, then type the result into the focused window.
All documentation
On this page
SigXor includes local voice input. Hold the hotkey, speak, and release. Recognition runs on your machine and the text is typed into the currently focused window. Audio never leaves the device.
Enable voice input#
- Open the tray menu → Settings → Voice & screenshot.
- Turn on Start voice service.
- On first use, download the SenseVoice model (about 230 MB) from Services & models.
- Make sure microphone access is allowed, focus the target input field, and start speaking.
Hotkey and interaction#
The default hotkey is Right Alt. You can change it to Left Alt, Right Ctrl, Left Ctrl or Caps Lock.
| Action | Behavior |
|---|---|
| Tap once | Start recording; tap again to stop and recognize |
| Press and hold, release | Record until release, then recognize |
| While recording | One capsule overlay with a waveform appears on each monitor; it never steals focus |
Languages and input methods#
- Languages: Chinese, English, Cantonese, Japanese, Korean and automatic detection.
- Clipboard paste (default): writes the result to the clipboard and pastes it, then restores the previous clipboard content.
- Keyboard simulation: types characters one by one with a configurable delay, for fields that do not accept paste.
The result is sent only to the focused window. Click the target field before speaking, or the text may land somewhere else.
Model management#
The speech model lives in Documents/FNSoftware/SigXor/Models/Speech/.
- View the model size, download it, delete it, or open its folder from Services & models.
- Reuse models from an older SigXor installation read-only, or move them manually.
- Deleting the model returns the app to the not-downloaded state; the next use downloads it again.
Privacy#
- SenseVoice and sherpa-onnx run entirely on-device.
- Audio is used only for the current recognition. Unless debug recording is enabled, it is not retained.
- The app contains no telemetry or analytics upload.
Known limits#
- Administrator windows: Windows UIPI blocks a normal-privilege process from typing into elevated windows. Run SigXor with the same privilege level if you need this.
- Hotkey conflicts: if Right Alt is captured by an IME or another tool, choose a different key.
- Accuracy: microphone quality, background noise, speaking speed and the selected language all matter; technical terms may need manual correction.
- First download: models cannot be fetched offline, but everything works offline once they are downloaded.
Troubleshooting#
If the hotkey does nothing or text is not typed into the target window:
- Confirm the voice service is running and the model is downloaded.
- Check for hotkey conflicts with other software.
- Check Windows microphone permissions and the default input device.
- Try switching from Clipboard paste to Keyboard simulation.
See Troubleshooting for more.