Fast, private speech input for Windows
Hold the left mouse button, or use a global hotkey, to transcribe speech directly into the active text field.
Mouse Dictation is a Windows-first desktop speech input tool for typing speech wherever your cursor is active.
 Recognition profiles, API or local offline mode, language selection, and model management. |
 Compact speaking overlay with the elapsed recording time. |
The application icon is shared across the Windows executable, installer, and tray assets.
Current version: 0.1.22
- Mouse press-and-hold or configurable global hotkeys
- Paid transcription APIs through OpenAI-compatible endpoints
- Local offline transcription with sherpa-onnx
- Realtime local transcription with a streaming Zipformer model
- Automatic translation after transcription
- Configurable recognition languages, models, themes, and profiles
- API keys stored in the operating system credential manager
- Windows Unicode text input without replacing the clipboard
- System tray support and optional launch at startup
Requirements:
- Node.js and npm
- Rust and the Tauri prerequisites for your operating system
Install dependencies and start the development app:
npm install
npm run tauri:devBuild the production application:
npm run tauri:buildConfigure an OpenAI-compatible API Base URL, API key, and transcription model in the Recognition tab. The default workflow starts recording after the mouse button is held for one second, then sends the recording to /audio/transcriptions when the button is released.
Remote API endpoints must use HTTPS. HTTP is allowed only for localhost or literal loopback IP addresses, including 127.0.0.1 and ::1, so local services such as Ollama continue to work. Plain HTTP on a LAN or public address is rejected before sending credentials or audio.
Realtime streaming depends on the provider and model. The paid API workflow currently uses final output after recording.
The request timeout scales with recording length (90 seconds minimum, 11 minutes maximum). Connection failures and temporary HTTP errors (429, 500, 502, 503, 504) can be retried once. Short numeric Retry-After instructions are honored; longer or date-based delays are returned as errors. Upload/response timeouts are not automatically retried because the provider may already have processed the audio.
On Windows, recognized text is entered through Unicode keyboard events without replacing clipboard contents. The target app must accept simulated text input and run with compatible permissions. Other platforms currently use clipboard-based paste.
The high-accuracy mode downloads and configures a SenseVoice int8 model. The realtime mode uses a streaming Zipformer model and inserts stable partial results while speaking.
Local mode processes audio on the device and supports manually selected model directories. Official model downloads are available from the sherpa-onnx ASR models page.
Automatic translation can be enabled after transcription. You can use a configured OpenAI-compatible translation provider or a local Ollama model. Translation sends the transcribed text to the selected provider; local Ollama keeps the request on the device.
- Local offline mode does not upload recordings.
- Paid API mode sends recordings to the API provider configured by the user.
- Automatic translation sends transcribed text to the selected translation provider.
- API keys are stored in the system credential manager rather than the normal settings file.
- The application does not include telemetry, advertising, or background analytics.
See PRIVACY.md for details.
On Windows, allow microphone access when requested. On macOS, microphone, Accessibility, and Input Monitoring permissions may be required for global mouse listening and simulated text input.
Before publishing a Windows release, test the application on a clean machine without the development environment installed:
- First launch and microphone permissions
- System tray and launch-at-startup behavior
- Mouse trigger and global hotkeys
- Overlay visibility and recording timer
- Local model download and deletion
- Paid API transcription
- Automatic translation
- Installer and uninstall flow
The Windows release currently uses an NSIS installer. Code signing is recommended for public distribution.
The source code is available under the Apache License 2.0. The Mouse Dictation name, brand, and icon are not granted as trademarks by that license.
Recording and global mouse listening run in Rust background threads. Local inference runs through sherpa-onnx; online performance depends mainly on network latency and the selected provider. Local performance depends on CPU speed, model size, and the first model load.