Talk. Hece types.
Hece is a free, open-source dictation app for Windows. Hold a key and speak; your words appear in whatever app you are using, transcribed on your own GPU.
Hold Ctrl Shift Space and start talking
Windows, 64-bit. Local mode needs an NVIDIA GPU; without one, use a cloud provider.
Private by default
Local Whisper runs on your machine. In local mode, audio never leaves your PC.
Speaks your language
Multilingual Whisper, plus a vocabulary that spells your names and terms the way you do.
Keeps your meaning
No phantom text from silence, and a cleanup step that is not allowed to drop a "not".
How it works
Your cursor is the destination.
Hece sits in the tray. When you hold the hotkey, it listens; when you let go, the text is pasted into the window you were typing in: a chat, an editor, a browser, a terminal.
- Hold Ctrl Shift SpaceOr switch to toggle mode: press once to start, once to stop.
- Speak naturallyA small orb in the corner shows that Hece is listening.
- Let goThe text lands where your cursor is, with one paste. Ctrl Shift Backspace cancels.
Quality
Say it once. Get what you meant.
Plain Whisper invents text in silence and mishears names. Hece adds checks between the model and your window.
your vocabulary
phantom text filtered
"can't" kept
Illustrative examples. Optional LLM cleanup only removes fillers and fixes punctuation; if its output drops a negation or changes length sharply, Hece pastes the raw text instead. Low-confidence results are flagged. The phantom-text filter knows the stock lines of 16 languages, including English, Spanish, German, French, Japanese and Chinese. The negation guard reads negations in 11, from English and Turkish to Polish and Russian; in others it still checks for empty output, added lines and big length changes.
Local or cloud
Your voice, your choice of engine.
Local
Whisper large-v3 turbo running on your NVIDIA GPU through CTranslate2. Fine-tuned models for single languages are one click away.
- Audio never leaves your PC
- The model loads when you need it and is released when idle
- Pinned and checksummed model download
- Other apps can share it through a local OpenAI-compatible API
Cloud
No GPU? Bring your own key for a provider you trust.
- OpenAI-compatible endpoints, Deepgram, Groq
- Keys are stored in the Windows Credential Manager
- Audio goes only to the provider you pick
- Switch back to local any time in the settings
No telemetry
Hece sends nothing about you; it only checks GitHub once a day for a new version, which you can turn off. Audio, transcripts and keys are never logged.
Out of clipboard history
Pasted text never enters Win+V history or the cloud clipboard.
History on your disk
Recent dictations stay local so you can paste one again. Turn it off or cap it in the settings.
Works where you paste
One paste chord, no key-by-key typing and no keyboard hooks.
Light when idle
A small tray app; the model only takes memory while you use it.
Open source, MIT
Read the code, build it yourself, or help shape the Linux and macOS versions.
FAQ
Questions, answered.
Is Hece free?
Yes. Hece is free and open source under the MIT license. There is no account and no subscription. In cloud mode you pay your provider directly, if they charge.
Which languages does it support?
The default local model is multilingual Whisper large-v3 turbo, and Hece can download fine-tuned models for single languages (Turkish today). Accuracy varies by language, as it does for Whisper itself. Cloud providers and any CTranslate2 Whisper model work too.
Do I need a GPU?
Local transcription needs an NVIDIA GPU with CUDA. Without one, pick a cloud provider (OpenAI-compatible, Deepgram or Groq) and add your key in the settings.
Does it work in every app?
It works anywhere you can paste: chat apps, editors, browsers, terminals. Windows blocks pasting into apps running as administrator; the text then stays in Hece's history so you can paste it yourself.
Windows shows a SmartScreen warning. Is that safe?
New, unsigned apps get this warning. Code signing is in progress. Until then you can check the source on GitHub and build it yourself.
What about macOS and Linux?
Hece is Windows-first. Linux and macOS are planned; the core is written to be portable, and help is welcome on GitHub.
How is Hece different from Windows voice typing (Win+H)?
Hece can run fully offline on your GPU, learns your vocabulary, filters phantom text, and lets you pick the model or provider. Win+H has no personal vocabulary.
