gujiguji v0.2.19

Highlights

• Keep Windows Microsoft Entra sessions across app restarts with WAM and an OS-protected persistent token cache.
• Refresh tokens silently and serialize interactive authentication so overlapping requests do not open duplicate sign-in windows.
• Preserve an active GPT transcription recording while required sign-in completes.
• Add explicit Azure account switching and clearer authentication/RBAC recovery messages.

Conditional Access, MFA, or tenant sign-in-frequency policies can still require an interactive sign-in.

Downloads

• Windows x64: VoiceInput.exe — self-contained single-file build.
• macOS Apple Silicon: gujiguji-mac.zip — Developer ID signed, Apple-notarized, and stapled.

SHA-256

• VoiceInput.exe: 5de962c5534786fb3778fda1cd35c0ff7dc3a970a32cc4c2e59eb0387812d96d
• gujiguji-mac.zip: 13829ce246c586792ca099ef948c0dd340319c145c0632e1bbb1c0bfab738326

Windows note: this executable follows the existing unsigned release path because no Authenticode certificate is configured; the installer and updater verify the GitHub-provided SHA-256 digest.

gujiguji v0.2.18

Highlights

• Press Esc to cancel an active dictation before transcription.
• Prevent Control plus keyboard shortcuts or mouse clicks from triggering dictation.
• Show update checking, download progress, verification, restart, failure, and retry directly in Settings on Windows and macOS.
• Improve input-state recovery and add focused regression coverage.

Downloads

• Windows x64: VoiceInput.exe, self-contained single-file build.
• macOS Apple Silicon: gujiguji-mac.zip, Developer ID signed, Apple-notarized, and stapled.

SHA-256

• VoiceInput.exe: 8121bfa12983325c186d7a5bfd90ecb48a62d3fa534dcdbfc4de77c3de2b6080
• gujiguji-mac.zip: ce22ecbff69d9007f72e2f7b774ca27d7502a5a038e2ba91b53167f91a97da39

Windows note: the executable follows the existing unsigned release path because no Authenticode certificate is configured; the updater verifies the GitHub asset digest.

gujiguji v0.2.17

What’s new

• macOS no longer starts local voice input while Microsoft Windows App is frontmost, allowing the activation key to pass through to the remote Windows session.
• Existing local dictation sessions can still stop or cancel normally after focus changes.
• Profile switching and menu actions remain available.

Downloads

VoiceInput.exe is the self-contained Windows x64 build. The macOS binary will be added after Apple notarization credentials are configured; the source release already contains the macOS behavior change.

The Windows executable is currently unsigned, matching previous releases. The supported installer and updater verify the GitHub-provided SHA-256 digest.

gujiguji v0.2.16

Windows hotfix

• Fixes Qwen3-ASR 1.7B model downloads failing with 403 Forbidden after redirecting to the ModelScope CDN.
• Sends a stable versioned User-Agent for both initial and resumed model downloads.
• Adds regression coverage for fresh downloads and Range-based resume requests.

Download VoiceInput.exe for Windows x64. It is self-contained and does not require a separate .NET installation.

The Windows executable is currently unsigned, matching previous releases. The supported installer and updater verify the GitHub-provided SHA-256 digest.

gujiguji v0.2.15

What’s new

• Qwen3-ASR 0.6B is now the default local model for new Windows and macOS users.
• Adds optional Qwen3-ASR 1.7B support for higher-accuracy local recognition.
• Improves model management, vocabulary prompts, settings layout, and cross-platform CI stability.

Windows download

Download VoiceInput.exe for Windows x64. It is self-contained and does not require a separate .NET installation.

The Windows executable is currently unsigned, matching previous releases. The supported installer and updater verify the GitHub-provided SHA-256 digest. The macOS asset will be added after Apple notarization is configured.

gujiguji v0.2.14

What's new

• Add two customizable input profiles with independent activation key, hold/toggle behavior, and top/bottom overlay position.
• Switch profiles with Alt+Shift+G or the tray menu; the last active profile is restored.
• Lead the README with a Copilot installation prompt and verify unsigned official releases with GitHub SHA-256 digests.

Self-contained Windows x64 build. No .NET runtime required. The executable is currently unsigned; Windows may request confirmation.

gujiguji v0.2.13

Unsigned self-contained Windows build. No .NET installation is required.

Highlights:

• Allow vocabulary learning to return more than 20 suggestions.
• Keep candidate validation and case-insensitive deduplication.
• Stage learned terms directly in Settings for review; Save applies them and Close discards them.
• Remove the modal vocabulary confirmation that could make the tray appear unresponsive.

SHA-256: EC66212AAB262BF12F9DECDCC9CA0E2B5937FBDF977D2107CE3F39BC6E1FDB3F

gujiguji v0.2.12

Unsigned self-contained Windows build. No .NET installation is required.

Highlights:

• Rebrand the product UI, icon, shortcuts, installer, and docs as gujiguji while preserving VoiceInput.exe and existing data paths.
• Prevent GPT transcription from emitting vocabulary on silent recordings.
• Remove the long Settings startup delay and simplify model selection.
• Show only engine-relevant model details.
• Add vocabulary suggestions from local correction history and allow loopback HTTP LLM endpoints.

SHA-256: 23E87E41FED7BB43F7B33BF2CDA6AA9D8A949EDB2F7066A9BCF46FDD2CBD60DF

VoiceInput v0.2.11

Unsigned self-contained Windows build (no .NET install needed).

• Add model-aware recognition vocabulary with comma, semicolon, or newline separators.
• Combine Windows, Azure Speech, GPT-4o Transcribe, and FunASR under Model Selection.
• Show local package contents, total size, download progress, and explicit cancellation.
• Keep active and pending model state visible, and require local downloads before saving.

VoiceInput v0.2.10

Unsigned self-contained Windows build (no .NET install needed).

• Default new users to FunASR SenseVoiceSmall local recognition.
• Let users choose hold-to-talk or toggle mode, then verify it in the onboarding window.
• Show model download size and progress, and preserve engines/models selected in full settings.
• Keep chorded shortcuts from advancing the onboarding practice state.