OSGKeyboard - AI voice input

New Features 1. Voice conversations support Picture-in-Picture (PiP) background retention (enabled by default); Dynamic Island remains an alternative option. Conversation ends immediately once PiP is disabled. 2. Added Tone Style Pack: users may select writing personas and customize prompts, with configurations synced via iCloud. 3. New playful polishing styles available: Straightener for Toxic Masculinity, Posturing Guide, Corporate Jargon, Tieba Legend, Xiaohongshu Girl Squad, and more. 4. Cloud recognition supports native streaming mode (Bailian / Volcano Engine / OpenAI). Text appears on screen in real time as you speak. 5. Added option to delete history records by single day (clear only records generated on the current day). 6. Streaming ASR providers are labeled "Streaming Recognition" in Settings. OpenAI cloud recognition defaults to streaming Whisper. 7. Mac now supports custom vocabulary with auto-generated recognition aliases, synced across devices via iCloud. 8. Local dictation on Mac is upgraded to Qwen3 MLX streaming recognition; text previews update while speaking. 9. Mac Settings now lets users select target languages for "Translate After Polishing". Experience Improvements 1. New installations use PiP background retention by default instead of Dynamic Island. 2. Default inactivity timeout for voice conversations adjusted to 5 minutes for better privacy and battery efficiency; users may switch between 1-minute / 5-minute timeouts in Settings. 3. Permission & privacy statements clarified: local device recognition is default. Audio files are only sent to your configured third-party cloud services when cloud mode is manually enabled. 4. Mac local recognition engine replaced with Qwen3 MLX; legacy offline Sherpa solution removed. 5. More stable model downloads for mainland China networks: mirror servers prioritized with resume download support. Manual download source selection available in Settings. 6. Buttons show "Next" instead of "Allow" on onboarding/home pages before system permission popups trigger. 7. Home status bar displays the active voice recognition provider and model name; incorrect display of polishing LLM fixed. 9. Tab icons for History and Lexicon on iPhone aligned with Mac and iPad designs. 10. Polishing style cards on Mac adopt responsive card grids that expand into denser layouts as the window widens. 11. Reorganized Settings page hierarchy: Transcription & daily functions prioritized; General and About sections moved to lower positions. 12. Streamlined UI for polishing style cards and setting summaries with fewer decorative elements and redundant text labels. 13. Unified card padding and rounded corner radius across Settings, style configuration and related pages. Bug Fixes 1. Resolved stuck final page issue after completing new user onboarding. 2. Welcome guide re-appears after app deletion and reinstallation. 3. Reduced trailing-word loss when releasing the microphone trigger; short utterances trigger full re-recognition as fallback. 4. Less chance of truncated opening speech during microphone warm-up upon record activation. 5. Fixed issue where fully transcribed text would get erased if the final recognition segment returns empty results. 6. Corrected text distortion of playful polishing styles when applied to short sentences.