How do I do voice dictation on Windows fully offline?
Windows ships with two built-in dictation paths: Win+H (the modern dictation overlay) and the legacy Speech Recognition control panel. Win+H requires an internet connection and sends audio to Microsoft's online speech service. The legacy Speech Recognition is local but its accuracy on natural speech, accents, and technical vocabulary is roughly a decade behind modern models like OpenAI's Whisper.
For genuine offline dictation on Windows with modern accuracy, install StarWhisper. It runs OpenAI's Whisper model entirely on your computer via the whisper.cpp engine, with no cloud component required. After the one-time download, you can disable your internet, fly on a plane, work in a Faraday-caged hospital room, and dictation still works. Press a hotkey, speak, and transcribed text types itself into the active window.
Five Whisper model sizes are bundled or downloadable: Tiny (75 MB) for fast CPU-only use, Base (142 MB), Small (466 MB), Medium (1.5 GB), and Large (2.9 GB). Each step up is roughly 2x to 4x slower. For everyday dictation across English and most Latin-script languages, Small gives the best speed-to-accuracy balance on Auto and is the right default for almost everyone. Bigger is not automatically better: Medium and Large hallucinate more on short clips. They are worth trying for long-form recordings (meetings, lectures) or non-Latin script languages such as Chinese, Japanese, Korean, Arabic, or Russian. NVIDIA GPUs with CUDA support (GTX 10 series and newer) accelerate transcription 5 to 10 times.
The free tier of StarWhisper covers 500 words per day forever, with no account or credit card required, and gives access to the Tiny, Base, and Small models. Pro at 10 USD per month removes the word limit and unlocks Medium and Large for the specific cases above (long-form audio, non-Latin scripts, batch file transcription).
Importantly, voice data is not stored or transmitted by default. There is no telemetry of audio content, no cloud upload, and no analytics tracking of what you dictate. This makes the app suitable for medical, legal, and other privacy-sensitive contexts where Win+H or any cloud dictation tool is not an option.
All FAQ topics