Skip to content

nana-yaw/hearme

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

16 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

HearMe

Downloads License

Private dictation that finds the model that hears your voice best — and never sends a byte anywhere.

🌍 hearme — one-page site & download

HearMe is a hardened, feature-extended fork of the excellent Handy by CJ Pais. Hold a hotkey anywhere on your desktop, speak, release — the transcript is typed at your cursor. Everything runs on-device.

Why this fork exists

Mainstream speech models are benchmarked on American and European voices. On West-African-accented English, the model topping the public leaderboards silently dropped 60–80% of the words in live tests, while another transcribed near-verbatim. You can't see that on a leaderboard — so HearMe measures it: the voice-setup wizard has you read three short phrases once, scores every model on your machine against your voice (word error rate + speed), and applies the winner in one click.

What the fork adds over upstream

  • Voice-setup wizard — benchmark all downloaded models on your own voice, apply the best
  • Local-only by construction — cloud LLM providers are removed from the code (not just disabled) and pruned from old settings on load; update phone-home off by default
  • Command mode — highlight text, hold Ctrl+Shift+Space, say "make this more formal": a local LLM (Ollama) rewrites it in place, reading the selection via the Accessibility API
  • Live typing (experimental) — words appear while you speak; the cleaned text re-settles on release
  • Spoken punctuation — "period" → "." deterministically, no LLM latency
  • Dictionary suggestions — recurring names from your history become one-click custom words
  • Per-app cleanup prompts, history search, mic-too-quiet warning, model activate toggle + launch preload, accent color, self-healing hotkey doctor

Full list: docs/FEATURES.md · release history: docs/CHANGELOG.md · design/internals: docs/ARCHITECTURE.md

Install

macOS (Apple Silicon, the tested platform): grab the .dmg from Releases, drag to Applications. The build is unsigned: right-click → Open on first launch, then grant Microphone, Accessibility, and Input Monitoring when prompted.

Windows (beta): ..._x64-setup.exe (Intel/AMD) or ..._arm64-setup.exe (ARM). Unsigned — SmartScreen will warn: "More info" → "Run anyway".

Linux (beta): .deb, .rpm, or .AppImage from the same release, in amd64 and arm64 flavors. Wayland support is limited upstream.

Windows and Linux builds are CI-built from the same source but not yet hand-tested on real hardware — issues welcome. Building from source: docs/INSTALL.md.

Don't trust — verify

lsof -i -a -p $(pgrep -x handy)   # while dictating: zero rows

The only network use is the one-time, SHA-256-pinned model download you trigger yourself.

Credits

  • Forked from cjpais/Handy (MIT) — thank you for the outstanding foundation.
  • Built end-to-end with Claude Fable 5 (Anthropic): research, code, adversarial reviews, docs, and the landing page.

License

MIT, same as upstream.

About

Private dictation that finds the model that hears YOUR voice best — 100% local, accent-aware, open source. macOS · Windows · Linux

Topics

Resources

License

Contributing

Stars

1 star

Watchers

0 watching

Forks

Packages

 
 
 

Contributors