Your voice. Your machine. Your text.
Suniye is open-source dictation for macOS. Speech recognition runs locally — nothing leaves your machine. Hold one key, speak, release.
If macOS says "App is damaged", run this command Suniye is not signed with an Apple Developer Certificate ($99/year), so the app is unsigned. macOS quarantines unsigned apps by default. This command removes the quarantine attribute.
xattr -rd com.apple.quarantine /Applications/Suniye.app
macOS 26+ · Alpha · MIT License
How it works
Hold your hotkey
Press and hold Fn/Globe — or any shortcut you configure. A floating indicator appears.
Speak naturally
Talk at your normal pace. Audio is captured locally through your microphone.
Text appears
Release the key. Your speech is transcribed on-device and inserted at your cursor.
Built for real work
Runs on your Mac
Speech models run entirely on-device using sherpa-onnx. Choose from a curated library — no internet needed for transcription. No server. No cloud.
Instant transcription
No network round-trip. Your voice becomes text in milliseconds, not seconds. The model runs in memory — always warm, always ready.
Works in any app
Text is inserted at your cursor via macOS Accessibility APIs. Mail, Slack, VS Code, Notes, your terminal — everywhere.
Optional LLM polish
Connect any OpenAI-compatible endpoint to clean up grammar, fix punctuation, or apply domain-specific vocabulary. Off by default.
One configurable shortcut
Default is Fn/Globe. Change it to any modifier combo you like. Hold to talk, release to transcribe. No menus, no buttons.
Open source, MIT licensed
Full source on GitHub. Read every line. No telemetry, no tracking, no surprises. Build from source if you want.
Every word you dictate could be training data.
Unless it never leaves your machine.
The problem
Cloud dictation services process your voice on remote servers. Your private notes, medical records, business plans — all passing through infrastructure you don't control. That audio can be stored, analyzed, and used to train models without your explicit consent.
Suniye's approach
Speech models run directly on your Mac — pick one from the built-in library. Voice is processed in memory, transcribed locally, and never touches a network. The only network calls Suniye makes are to check for app updates and download models.
When you enable the optional LLM cleanup, transcribed text (not audio) is sent to the provider you choose. This is opt-in and disabled by default.
Try Suniye
Download the latest release or build from source. Requires macOS 14 (Sonoma) or later. Pick a speech model from the built-in library on first launch.
If macOS says "App is damaged", run this command Suniye is not signed with an Apple Developer Certificate ($99/year), so the app is unsigned. macOS quarantines unsigned apps by default. This command removes the quarantine attribute.
xattr -rd com.apple.quarantine /Applications/Suniye.app
Alpha — expect rough edges