* feat(text): Add n-gram matching for multi-word custom word correction
Improves custom word matching to handle speech artifacts where a single
word gets transcribed as multiple words (e.g., "Charge B" -> "ChargeBee",
"Chat G P T" -> "ChatGPT").
- Implements greedy n-gram matching (3 to 1 words) for better accuracy
- Adds length-based filtering (25% max difference) to prevent over-matching
- Extracts matching logic into reusable find_best_match function
- Preserves punctuation and case from original transcription
- Adds comprehensive test coverage for n-gram scenarios
* minor tweak
* format
---------
Co-authored-by: CJ Pais <cj@cjpais.com>
* fix: run Rust tests in CI
Fixes#670
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* fix: run Rust tests in CI
Adds a separate test.yml workflow to run existing Rust unit tests.
Fixes#670
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* refactor: use mock TranscriptionManager in CI tests
Swap to a mock adapter during CI to avoid compiling whisper/Vulkan.
The mock has the same interface but no heavy dependencies.
Existing tests don't exercise transcription code, so this is safe.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
* Add Nix flake for NixOS support
Provides:
- handy-appimage: AppImage-based package (default)
- devShell: Development environment for building from source
* fix(nix): address PR feedback for flake.nix
- Remove aarch64-linux from supportedSystems (ARM64 builds not produced)
- Read version dynamically from Cargo.toml instead of hardcoding
This enables automated updates during GitHub releases
* chore(nix): add GitHub Actions workflow to auto-update AppImage hash
Adds a separate workflow triggered on release publication that:
- Fetches the released AppImage
- Computes the SRI hash using nix-prefetch-url
- Updates flake.nix with the new hash
- Commits and pushes changes directly
Includes retry logic for timing issues and defensive checks to verify
the hash was actually updated and skip commits if unchanged.
* Update flake.nix hash to resolve mismatch
Co-authored-by: pinage404 <pinage404@gmail.com>
* Update flake.nix to install dependencies automatically
Co-authored-by: pinage404 <pinage404@gmail.com>
* fix: update AppImage hash for v0.7.0
* modified description
---------
Co-authored-by: pinage404 <pinage404@gmail.com>
Co-authored-by: CJ Pais <cj@cjpais.com>
* init attempt at new kb
* 0.1.1 handy-keys
* format
* cleanup
* support for capslock/shiftlock?
* move kb to experimental
* use hook
* fix macos bug
* attempt mouse buttons as well
* Add Portuguese translations for permission error messages
Add error message translations for the permission check flow in the Portuguese locale, covering both check failures and request failures.
* Fix history audio playback on Linux
WebKitGTK cannot play audio files via the asset:// protocol (or at least
not reliably), returning MEDIA_ERR_SRC_NOT_SUPPORTED. Additionally,
convertFileSrc was double- encoding the path (/ became %2F).
Changed to read audio files directly using the fs plugin and create
blob URLs instead. Also added APPDATA to fs:scope permissions to allow
reading recordings from the app data directory.
* Fix audio playback on macOS by using asset protocol on non-Linux
The blob URL approach for audio playback works on Linux but breaks
macOS.
This reverts the blob change to use convertFileSrc with the asset
protocol on macOS/Windows while keeping the blob approach for Linux
where it's needed. Also adds proper cleanup of blob URLs to prevent
memory leaks when history entries are unmounted.
* refactor: use Immer for immutable state updates
Replace Set/Map with Record types and use Immer's produce() for
immutable state updates. This fixes mutation bugs where .add()/.set()
were mutating state before copying (e.g., `new Set(prev.add(id))`).
Changes:
- Add immer dependency
- Convert Set<string> to Record<string, true> (sparse hash set pattern)
- Convert Map<K,V> to Record<K,V>
- Use produce() for all state mutations
- Update prop types in child components
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* move deps back to where they started
---------
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
Co-authored-by: CJ Pais <cj@cjpais.com>
* feat: add automatic filler word removal from transcriptions
Add filter_transcription_output() that removes filler words (uh, um, hmm,
etc.) and hallucination patterns ([AUDIO], (pause), <tag>...</tag>) from
transcriptions. Inspired by VoiceInk's approach.
- Add regex-based filter in audio_toolkit/text.rs
- Integrate into transcription pipeline after custom word correction
- Add comprehensive tests for filler word removal
- Add regex crate dependency
* feat: collapse repeated 1-2 letter stutters in transcriptions
Add collapse_stutters() to reduce model hallucination artifacts like
"wh wh wh wh wh why" -> "wh why" or "I I I I think" -> "I think".
Collapses any 1-2 letter word that repeats 3+ times consecutively
to a single instance (case-insensitive matching).
* remove bracket/xml stuff
* format
---------
Co-authored-by: CJ Pais <cj@cjpais.com>
* Add permission step to onboarding
* translations
* minor ui
* Update AccessibilityOnboarding.tsx
* claude suggestions
---------
Co-authored-by: CJ Pais <cj@cjpais.com>
- hide model dropdown when apple intelligence is selected
- move error state into provider section (only shows when unavailable)
- remove redundant apple intelligence infobox
- add reusable Alert component with variants (error, warning, info, success)
- Alert supports contained prop for use inside settings groups
* Fix race condition when toggling transcription via SIGUSR2
The toggle state was being set to false immediately when stop was called,
but the actual transcription runs asynchronously. If another signal arrived
before transcription completed, a new recording could start while the old
transcription was still running, causing the old text to paste unexpectedly.
Now the toggle state stays true until the async transcription task completes.
* remove import
---------
Co-authored-by: CJ Pais <cj@cjpais.com>
Add ydotool as a fallback option for direct text input and key
combinations on Linux, after wtype/dotool (Wayland) and xdotool (X11).
ydotool uses uinput and works on both Wayland and X11, making it a
useful alternative when other tools are unavailable.
Requirements:
- ydotool installed
- ydotoold daemon running
- User in input group with uinput permissions
* feat: add Moonshine Base speech recognition model
Add support for the Moonshine Base ASR model, which offers:
- 5x faster transcription than Whisper
- Better accent recognition
- ~240MB model size (encoder + decoder + tokenizer)
Changes:
- Update transcribe-rs to v0.2 with moonshine feature
- Add EngineType::Moonshine variant
- Add Moonshine Base model definition (400MB, 80% accuracy, 90% speed)
- Add model loading and transcription logic
- Add i18n translation for Moonshine Base
Note: Model files must be hosted at blob.handy.computer/moonshine-base.tar.gz
Expected structure: encoder_model.onnx, decoder_model_merged.onnx, tokenizer.json
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* Update moonshine parameters.
* add translations
---------
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
Co-authored-by: CJ Pais <cj@cjpais.com>
* Add `select-none` and `cursor-default` to root to prevent selection and
hover on most items
* Lint
* refactor paths a bit
* format
---------
Co-authored-by: CJ Pais <cj@cjpais.com>
defer apple intelligence availability check from app initialization
to when the user actually tries to use the feature. accessing
SystemLanguageModel.default during early app startup causes SIGABRT
on macOS 26.x (tahoe) beta.
- remove availability check from default_post_process_providers()
- always include apple intelligence provider on macos arm64
- add checkAppleIntelligenceAvailable command for lazy checking
- show error message if apple intelligence unavailable when selected
- clear error on any dropdown selection for better UX
* fix: unload model on cancel when immediate unload enabled
adds model unload check to cancel_current_operation() to match
the behavior in transcribe(). previously, cancelling with escape
would leave the model loaded even with "unload immediately" enabled.
fixes#479
* also fix model not unloading if there is empty transcription
* refactor to util function
---------
Co-authored-by: CJ Pais <cj@cjpais.com>
* fix: replace async-openai library with post request to support custom providers
* move models call to llm_client too
* tweak post processing ui and add groq and cerebras
---------
Co-authored-by: CJ Pais <cj@cjpais.com>
* Add Russian language
* Add Russian language
* Clean up hidden Unicode characters in Russian translation
* add languages.ts
* add some missing translations
---------
Co-authored-by: CJ Pais <cj@cjpais.com>
