Cloud service and voice-input improvements
- Improves automatic switching between cloud services.
- Reduces false transcripts from recordings without speech.
- Adds a System language option for new Shuo installations.
- Improves the stability of live audio capture from USB microphones.
Expanded cloud service support
Adds Groq, SiliconFlow, and custom OpenAI-compatible endpoint support, with improved cloud model and API key configuration.
Custom push-to-talk shortcuts
Shuo 1.2.3 adds a Custom push-to-talk shortcut option in onboarding and Settings.
Dock, update-check, and Floating Bar polish
Dock and menu-bar behavior
- Shuo now stays menu-bar-only by default, with a Settings toggle for users who prefer to show the Dock icon.
Cleaner update checks
- Check for Updates now shows “You're up to date” directly in Settings when no update is available, and the message is localized.
Better mixed-language punctuation
- Mixed Chinese/English transcripts now preserve model-provided sentence-ending punctuation instead of rewriting it based only on the final token's language.
Floating Bar polish
- The compact Floating Bar now keeps the arrow cursor when hovering over the logo-only state.
A more reliable menu bar app launch
- Fixed a launch crash that could occur before Shuo's menu bar item appeared.
Cloud Gemini and a more dependable Floating Bar
Cloud Gemini API support
- Added cloud Gemini API support using Gemini 3.1 Flash-Lite.
More dependable Floating Bar results
- Reduced cases where the Floating Bar disappears immediately after a result is inserted.
Clearer local-model choices and more dependable input
A clearer local-model choice
- Added SenseVoice Small, a fully local option tuned for Chinese, Japanese, and mixed speech. It does not support prompt context, custom vocabulary, or project-terminology hints.
- Local downloads are now intentionally limited to three understandable choices: SenseVoice Small, Whisper Small, and Whisper Large Turbo. Shuo recommends one based on your selected languages and Mac.
- Each model now explains its language range, size, and support for context hints and custom vocabulary before you download it.
More dependable input and paste
- System Default is now the default audio-input choice, making external microphones and AirPods follow macOS routing more predictably.
- Improved microphone startup and audio-format changes so a Bluetooth or external microphone is less likely to lose the opening of an utterance.
A smaller daily surface with a stronger foundation.
Shuo centers the default experience on “hold—speak—release.” Settings holds everyday choices, Advanced maps every setting onto a seven-stage pipeline, and About collects permissions, updates, support, and local-data tools.
Input and recovery
- Choose Right Command or Right Option directly from Home; the remaining everyday controls are grouped in Settings.
- Successful dictation always inserts immediately; the redundant auto-insert setting has been removed.
- The optional Floating Bar stays as a thin idle indicator, expands after insertion, and lets you edit and replace the latest result.
- The Floating Bar is draggable, remembers its position across displays, and cannot be manually resized. Its right-click menu contains only Hide Floating Bar, Open Shuo, and Quit Shuo.
- The latest-result menu includes Copy, Replace, Play, and Redo with compact labels.
- The menu-bar panel now shows Settings instead of Preferred Terms and opens the Settings page directly.
- Replace and voice edits reactivate the original app and rewrite only the changed suffix while the recent insertion, app process, and interaction guard still match; otherwise Shuo copies the complete correction.
- Clipboard snapshots preserve representations, never overwrite newer user content, and bound waits on slow owners.
Quiet speech and local models
- Whisper Mode replaces its fixed gate with per-recording noise/speech measurement and bounded gain on a temporary transcription copy.
- Whisper Mode is available in Settings and the menu-bar panel; manual tuning lives under Advanced → Audio processing.
- whisper.cpp downloads show progress and bytes, check free space, support cancellation, and verify exact size and SHA-256.
- The direct app bundles a universal whisper.cpp runtime; Local setup only downloads a model and no longer requires Homebrew.
- Local transcription processes support cancellation and a ten-minute timeout.
Terms from the work in front of you
- Preferred Terms bias Local, OpenAI-compatible, and ElevenLabs Scribe transcription.
- Opt-in Project Vocabulary read-only indexes linked local names, manifests, paths, symbols, and documentation, retaining only the 60 highest-priority candidates for each project.
- Each transcription has a total vocabulary budget of at most 60 terms and 900 characters, shared by built-in terms, Preferred Terms, eligible correction-learned terms, project terms, and preset terms.
- Cloud providers that support terminology receive only selected spelling hints, never project source or paths; the current Alibaba adapter sends no vocabulary hints.
Corrections and data safety
- Edited quick Copy/Replace, History saves, and voice edits capture complete before/after text linked to History and audio when possible.
- Learning from Manual Corrections and every individual pattern are off by default. Once enabled, the learning method is mutually exclusive: Replacement applies only on this Mac after three distinct trusted sessions agree with 100% confidence and no conflict, and sends no correction-learned term; Cloud AI sends only corrected wording B to a hint-capable cloud transcription provider after at least two observations, 75% confidence, and a lead of two over the runner-up, and never rewrites locally.
- Advanced → Human correction → Learning history retains complete edits separately from legacy records and can export or clear them; legacy records never re-enter transcription.
- History and Metrics preserve corrupt files, recover from readable backups, and surface unrecoverable errors.
- Metrics is now a primary destination below History rather than a diagnostic page.
- Audio linked to retained History items stays local with raw, processed, and final text until the item is deleted; explicit edits show a local-save confirmation.
First run, cloud, and updates
- One-screen onboarding covers the shortcut, permissions, and Local or OpenAI-compatible transcription.
- ElevenLabs and Alibaba adapters remain optional Beta profiles until they pass real-service release QA.
- Authenticated OpenAI-compatible `/models` discovery, 24-hour caching, task-specific filtering, and Automatic/Fixed modes.
- The direct build uses a signed Sparkle appcast; the App Store-compatible target remains buildable but is not a 1.0.0 release channel.
- 1.0.0 publishes the GPL-3.0 source alongside the release; the isolated Shuo Community build does not use official data, Keychain identities, or the production update channel.
- Before installing an update, Shuo checks for the shared app running in another macOS account and waits without terminating that process.
- About centralizes Privacy, Release Notes, Uninstall & Data, permissions, feedback, diagnostics, settings export, data folders, and API-key removal; Check for Updates lives under Settings → Application.
- Advanced opens with no stage selected and maps the complete settings set onto a seven-stage pipeline. Search sits below that overview, and selecting a stage reveals all related controls.
Release status
Shuo 1.3.0 is released as a direct download for macOS. Every official ZIP and DMG corresponds to its public source tag, checksum, and signed update feed.