Skip to content

Releases

Hece v0.1.0

Hece v0.1.0 Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 03 Nov 06:34
v0.1.0
29287f6
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

The first public release of Hece: background voice dictation for Windows. Hold
Ctrl+Shift+Space, speak, and the text lands in the window you are typing in.

Install

  1. Download Hece_0.1.0_x64-setup.exe below and run it (per-user, no administrator rights).
  2. Hece Settings opens. Download a model (multilingual Whisper large-v3 turbo by default) if you
    have an NVIDIA GPU, or add a cloud provider key (OpenAI-compatible such as Groq, or Deepgram).
  3. Start Hece from the Start menu and hold Ctrl+Shift+Space to dictate.

Unsigned build. This release is not code-signed yet, so Windows SmartScreen may show
"Windows protected your PC". Choose More info → Run anyway. You can verify the file first:
compare it with SHA256SUMS.txt, or run
gh attestation verify Hece_0.1.0_x64-setup.exe --repo altintasutku/hece to check that it was
built by this repository's GitHub Actions workflow. Signing is in progress (#49).

What's in it

  • Push-to-talk and toggle modes with shortcuts you can change in Settings, a recording orb that
    reacts to your voice, short start/stop sounds (can be turned off), tray menu.
  • Local Whisper on your NVIDIA GPU (CTranslate2), fully offline once downloaded. The default model
    is multilingual Whisper large-v3 turbo; a fine-tuned Turkish model is one click away. The CUDA
    runtime (~430 MB) is downloaded on demand from NVIDIA. Models and runtime are pinned and
    checksummed. Supported GPUs: RTX 20 series (Turing) and newer.
  • Cloud providers with presets: Groq, OpenAI, Deepgram, or any OpenAI-compatible server. Keys are
    stored in the Windows Credential Manager. See docs/cloud-providers.md.
  • Quality layers: a vocabulary you edit in Settings, a filter for text Whisper invents in silence
    (16 languages), a low-confidence signal, and optional LLM cleanup with a guard that keeps
    negations (11 languages).
  • One paste chord into the focused window; transcripts stay out of Win+V clipboard history.
  • Local dictation history with re-paste, and an optional local OpenAI-compatible API.
  • English interface, Turkish when Windows is set to Turkish.
  • No telemetry. Hece checks GitHub once a day for a new version (turn it off in Settings → About).
    Audio, transcripts and API keys are never logged.

Known limits

  • Windows x64 only for now, tested on Windows 11. Linux and macOS are planned (help wanted).
  • Local transcription needs an NVIDIA GPU (RTX 20 series or newer); otherwise use a cloud provider.
    Only RTX 40/50 cards were tested before this release.
  • Pasting into apps running as administrator is blocked by Windows; the text stays in the history.

Full list of changes: see CHANGELOG.md.

Assets 4
Loading