- Published on
Show HN: Hushkey β Silent dictation hotkey for macOS
- Authors
- Name
- DeepHorus
GitHub developer amirrsyed has released Hushkey, a proof-of-concept silent dictation tool for macOS that turns mouth movements into text at the cursor. Users hold a hotkey, mouth words without sound, release the key, and text appears in any app. Unlike audio-based tools such as Wispr Flow, Hushkey reads lips, jaw, and throat motion through the webcam.
The project runs entirely on the Mac, with no audio capture and no video leaving the device. It requires macOS 14 or later, recommends Apple Silicon, and needs about 2.5 GB of disk space. A one-time model server download is roughly 1.5 GB, and users must grant Camera and Accessibility permissions.
Hushkey is distributed as an unnotarized Hushkey.dmg and can also be installed from source via GitHub. Unsigned builds may require re-adding Accessibility access after updates, and if typing is blocked the app copies text to the clipboard for manual pasting. The camera turns off 30 seconds after the last dictation unless auto-detect is enabled.
Accuracy depends on the pretrained lip reader and on personalization. Users can train adapters on their own enrollment clips, run word-error-rate fine-tuning, and optionally add local LLM cleanup through Ollama with qwen3:4b. The developer describes the pipeline as end-to-end but cautions that results vary.
For best results, Hushkey advises facing the camera with mouth and neck visible, using even lighting, exaggerating slightly, and dictating phrases rather than single words. Because it avoids audio, the tool targets quiet or crowded environments where speaking is impractical. Its local-only design also distinguishes it from cloud dictation services, though webcam framing limits mobility.
As a proof of concept, Hushkey is not yet a polished consumer product. Its release signals growing experimentation with multimodal, on-device input for privacy-sensitive workflows. Future versions may improve accuracy through better models, broader training data, and deeper personalization.
