Local Voice v0.2.2

Small but important UX fix, directly addressing the confusion that caused the v0.2.0 double-launch bug.

What's new

• Toast notifications for "Starting up..." and "Ready — hold F9 to dictate" — no more wondering whether anything happened after clicking the shortcut.
• A visible "already running" dialog if you launch a second time (e.g. by double-clicking impatiently) instead of silently doing nothing.
• Note: the tray icon itself was always there the whole time the app runs — Windows just tends to tuck new tray icons into the hidden-icons overflow area (the ^ arrow next to the clock) by default. Worth dragging it out to the visible taskbar area if you want it always in view.

Install / upgrade

Download LocalVoice-Setup-0.2.2.exe below and run it once. No admin rights needed.

Licensed under Apache 2.0.

Local Voice v0.2.1

Bug-fix release — fixes cloud cleanup actually working, and a real double-launch bug.

Fixes since v0.2.0

• Cloud cleanup was broken. The Anthropic API call used parameters (temperature, then effort) that this SDK/model combination rejects. Fixed to call with no generation-control parameter. Verified live: correctly resolves mistranscribed names using context, fixes grammar, leaves everything else untouched.
• Duplicate-instance bug from double-launching. Clicking the Start Menu shortcut a second time while the first launch was still loading the model (~3-5s) created two competing hotkey listeners, causing exactly the "F9 doesn't work" symptom. Added a proper single-instance guard — a second launch attempt now detects the first and exits immediately instead of racing it.

Upgrading from v0.2.0

Same installer flow — run it once, wait for it to finish. If the app seems unresponsive right after launch, that's the model loading (a few seconds) — wait rather than clicking the shortcut again.

Install

Download LocalVoice-Setup-0.2.1.exe below and run it once. No admin rights needed.

Licensed under Apache 2.0.

Local Voice v0.2.0

New feature release.

What's new

• Optional cloud grammar-cleanup backend (Claude Haiku 4.5 via the Anthropic API). The local cleanup model sometimes can't fix a wrong character/homophone even with context, simply because it's a small model — a frontier model can. This only ever sends already-transcribed text (never audio) and is off by default; see the README's "Cloud cleanup" section to enable it. Requires your own ANTHROPIC_API_KEY, never stored by the app. The model id is pinned to a specific dated snapshot and never silently auto-upgrades.

Upgrading from v0.1.x

Same installer flow — run it once, wait for it to finish. Existing settings/history/memory are untouched (nothing changed in %LOCALAPPDATA%\LocalVoice).

Also included

Everything from v0.1.1 (installer idempotency fix, the silent model-download-corruption fix, the dropped-speech-on-long-clips fix) — see docs/DECISIONS_AND_RISKS.md for full detail on all of these.

Install

Download LocalVoice-Setup-0.2.0.exe below and run it once. No admin rights needed.

Licensed under Apache 2.0.

Local Voice v0.1.1

Bug-fix release. If you installed v0.1.0, this fixes three real issues found during real-world testing — upgrading is strongly recommended.

Fixes since v0.1.0

• Installer: "Unable to execute file ...pythonw.exe" on re-run. The environment setup step wasn't idempotent — re-running the installer (or its own post-install auto-launch racing a second run) could try to overwrite files the just-launched app was still holding open. Fixed: it now detects an existing working environment and skips straight to a fast dependency check, and stops any running instance from that install directory first.
• Silent model-download corruption causing garbled transcription. huggingface_hub's newer default download path could silently fail to materialize the actual model weights (a ~3.5 MB result instead of ~3 GB), while reporting success with no error anywhere. This produced plausible-looking but fabricated transcription output that was easy to mistake for a speech-recognition quality issue. Fixed by forcing the classic HTTP download path.
• Long dictations only transcribing the ending. A decoding option that conditions each segment on previously-decoded text (Whisper's default) could let an early misstep cascade into dropped or corrupted later content. Disabled it.

See docs/DECISIONS_AND_RISKS.md for full technical detail on each of these.

Install

Download LocalVoice-Setup-0.1.1.exe below and run it once, waiting for it to fully finish before doing anything else. No admin rights needed. If upgrading from v0.1.0, remove the old install directory first (%LOCALAPPDATA%\Programs\LocalVoice) if it's in a bad state.

Licensed under Apache 2.0.

Local Voice v0.1.0

First release: a local, offline-first Windows dictation app for English, Mandarin, and mixed Chinese-English speech. Hold a hotkey, speak, release — GPU-based transcription, no cloud.

What's in this release

• Push-to-talk hotkey capture; auto-restarts into a new chunk if you hit the recording cap without releasing, so a forgotten key doesn't silently lose audio
• Local GPU transcription (faster-whisper, large-v3), auto-detects English/Mandarin/mixed and preserves code-switching (never translates)
• Clipboard delivery, with optional one-shot auto-paste (SendInput, focus-rechecked, never retried)
• Optional local grammar-cleanup pass with session context and a small persisted cross-session memory of names/recurring terms, to help disambiguate homophone errors
• Vocabulary hints for names/terms the model otherwise misses
• System tray app — no console window needed to run or stop it
• Update checking (right-click the tray icon)

Known limitations

• No settings UI yet — edit %LOCALAPPDATA%\LocalVoice\settings.json by hand (see README)
• The optional grammar-cleanup feature needs Ollama installed separately, with a model such as qwen3.5:4b pulled
• Rare/obscure vocabulary (e.g. classical poetry) can still transcribe with plausible-sounding wrong characters — see docs/DECISIONS_AND_RISKS.md for the details and what was tried
• The installer's first-run setup (downloads ~2 GB of CUDA/ML packages) has been validated to compile and run correctly on a machine that already has the dependencies, but not end-to-end on a genuinely clean machine

Install

Download LocalVoice-Setup-0.1.0.exe below and run it. No admin rights needed — installs per-user with a Start Menu entry and optional desktop shortcut. First run takes a few minutes to set up the Python environment and download dependencies.

Licensed under Apache 2.0.