Skip to content
ruslanvakhitovPublic
forked from anomalyco/hex

About

Local contextual voice dictation for macOS. Independent Hex fork with on-device screen OCR and vocabulary per app.

Resources

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

 
 

Latest commit

 

History

1,362 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Volna

Local voice dictation with vocabulary that understands the app you are using. An independently branded macOS fork of Hex.

  • Option: hold, speak, release. Dictation reuses the active app’s saved text context.
  • Right Command: hold, speak, release. Volna takes one screenshot of the active app’s display and reads its Russian/English text locally while you speak.
  • Per-app context: screen text remembered in Telegram stays with Telegram. Plain text helps correct names, acronyms and technical vocabulary after speech recognition.
  • Local processing: Apple Vision OCR and Qwen3.5-0.8B through a private, cloud-disabled Ollama runtime. No API key. History keeps the screenshot, extracted text, reference text and the before/after correction locally for each retained dictation. Screenshots follow History retention and a 128 MiB cap. Saved app text expires after 30 days.
  • Inspectable results: History shows OCR errors, deadlines, dictionary-only fallback and unchanged results, as well as corrections. Existing records cannot recover screenshots from earlier versions. Turning History Off stops new evidence.
  • New identity: wave icon, sea-glass theme, voice ribbon, glass-ripple sounds.

This version targets Apple silicon on macOS. Give Volna its own Microphone, Accessibility, and Screen Recording permissions. Upstream Hex’s files and grants are independent. Once Volna’s dictation setup is ready, it asks the old Hex to quit so they do not both answer Option. It never deletes the old application.

Build on a Mac

Requires Rust 1.97 or later, Xcode 26 (or macOS 26.1 Command Line Tools), and Ollama on the build Mac. The local correction model is about 1.3 GB.

scripts/setup-volna.sh
scripts/build-volna.sh
scripts/check-volna.sh
open dist/Volna.app

Output: dist/Volna.app, ad-hoc signed for local use. The builder includes the installed Ollama runtime and its licenses. The model lives separately in ~/Library/Application Support/volna/context-models. This is a local build, without Developer ID notarization or upstream automatic updates.

Settings → General includes the contextual shortcut, screen access, learning, local correction and inspectable app context. Forgetting an app also invalidates learning from a screenshot already in flight. Shared, manually edited vocabulary and the upstream speech-model picker remain available. Context is stored as up to four recent plain-text screen excerpts per app, without keyword ranking. Keyword-only stores from earlier builds remain on disk but are not used as text context; the next Right Command dictation starts the new app memory.

# Read status or inspect learned terms; these do not take a screenshot.
dist/Volna.app/Contents/MacOS/volna context status
dist/Volna.app/Contents/MacOS/volna context dictionaries
# Test the running app without audio, paste or retained History.
# Keep the target app in front; learning requires --learn.
dist/Volna.app/Contents/MacOS/volna context probe --transcript "открой кодекс"
# Test OCR on your own image without learning.
dist/Volna.app/Contents/MacOS/volna context ocr sample.png
# Test local correction on supplied text, without learning.
dist/Volna.app/Contents/MacOS/volna context rewrite \
  'Посмотрим кодекс в 15 часов' --screen 'Codex — developer tools'

See screen-context behavior and verification. The rest of the inherited capabilities are documented under docs/features/. The CLT fallback cannot build Apple’s FoundationModels macro plugin; optional meeting summaries using Apple Intelligence are unavailable in that build. Dictation, screen OCR and local Qwen correction do not use FoundationModels.

License

MIT. Upstream copyright and license are retained in LICENSE. See THIRD_PARTY_NOTICES.md and VOLNA_NOTICES.md for third-party components and models.

About

Local contextual voice dictation for macOS. Independent Hex fork with on-device screen OCR and vocabulary per app.

Resources

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages