New Features 1. Live Captions (beta): real-time transcription shown above the record button while recording (fixed height, no layout jumps), toggle in Settings 2. New editing mode on the detail page: (1) Find: case-sensitive / whole-word matching, previous / next navigation with auto-centering, match counter, replace (2) Highlight: select text, four-color palette (yellow / green / red / blue), automatic dark text on bright backgrounds, blue selection background with underline preview, stackable (3) Edit: paragraph-by-paragraph editing, cancel to roll back, save on finish; keyboard avoids focus, disabled during find / highlight 3. Transcript export: 6 formats, new PDF and DOCX, plus SRT, CSV, Markdown and TXT; optional speakers / timestamps / highlights / audio; entry in the more menu at the top-right of the preview page 4. Detail-page player: waveform card, progress, mute, jump within a segment 5. Discard recording anytime: "Discard" button at the top-right of the home screen, with a unified confirmation flow 6. Save status indicator: persistent "Saved at 00:01:51" on the home screen, with a "Save delayed..." warning on failure 7. Paywall policy and daily quota: after the trial ends you can still record 1 recording per day for free (never hard-blocked); the paywall only auto-triggers when you tap record after using up the daily quota, at most once per day; new bundle product added 8. History enhancements: inline export / rename, batch export after filtering (SRT / CSV), multi-select bulk delete; "History" tab renamed to "Recordings" 9. Others: greeting popup (with "Don't show today"), more menu, collapsible cards, horizontal confirm dialog, local analytics, SenseVoice engine 10. Long-recording performance: paragraph batch rendering in the detail page (first 40 paragraphs, load more on scroll, auto-expand on find / edit jumps) Changes 1. Recording interaction reworked: single tap to start / pause (with pause badge), double tap now asks for confirmation first (save or discard); complex animations removed; marker button style and position adjusted 2. Trial: reduced from 10 days to 7 days, plus a one-time +7 day extension per user 3. Recording recovery: no more "delete / continue" dialog, now auto-recovers and re-transcribes pending items 4. Streaming transcription and recording service: stop flow and timeout fallback reworked, persistence logic optimized 5. Localization in 5 languages: about 80 new / adjusted strings (Chinese / English / Japanese / Korean / Cantonese) Removals 1. "Restore Purchases" entry removed from Settings (moved to the Paywall screen, not actually deleted - no review risk) 2. "Clear All Data" entry removed from Settings (the "Clear" action in History remains and can still clear all recordings) 3. Forced paywall on launch (lock mode when the trial expired) replaced with a soft daily-quota gate 4. Paywall repeat-display restriction logic simplified 5. Debug-related reset features removed 6. Complex recording screen animations removed

1. Recognition Engines (1.5) Three engines: whisper-small / whisper-turbo / sense-voice, each declaring backend, best-fit languages, and minimum RAM. Selection: user choice > best for locale > RAM-eligible fallback, with reasoning logged. SenseVoice via sherpa-onnx: Chinese, English, Japanese, Korean, Cantonese; full download / import / delete lifecycle (938 MB). Silero VAD segmentation: whole-file voice-activity detection then per-segment decoding replaces fixed slicing — much better sentence boundaries on long recordings. RAM gate: engines filtered by available memory, so low-memory devices never load a model the OS would kill. Model management: "Model Engine" card with multi-mirror resumable downloads, cancellation, swipe/long-press delete, and a star on the active engine. Preference persisted per locale. 2. Speaker Diarization Replaced the legacy MFCC/ECAPA module with a sherpa-onnx pipeline (pyannote segmentation + CAM++ embeddings); cluster centroids map onto saved profiles, identifying the same person across recordings. Sensitivity: high / medium / low replaces a hardcoded threshold. 1.5 fix: the manual-assignment path was short-circuiting automatic diarization, leaving Chinese transcripts as one block; it now applies only when the user actually switched speakers. 3. Recording Reliability (1.5) Failure salvage: when stop fails because the OS reclaimed the audio session, persisted data is kept — header repaired, duration derived from data size, saved as "interrupted". Previously this orphaned the recording and falsely reported "too short". WAV + VAD fixes: header probe widened for AVAudioRecorder's 4096-byte header, with repair on stop and playback; fixed an uninitialized VAD window that stalled long Chinese recordings. Periodic persistence (15/30/60 s) plus leftover-recording recovery at launch. Streaming guard: non-streaming engines no longer arm the slice timer, fixing a recurring 30-second header-parse failure, "0 frames" errors, and the record button locking up after an engine switch. Background recording: continues when locked or backgrounded; auto-pauses when a call or another app takes the mic, then resumes. Added monitoring for route changes, media-services reset, memory pressure, and power state. 4. Localization Added complete Japanese, Korean, and Cantonese translations — five languages, all 295 keys. Cantonese uses Traditional characters and real Cantonese vocabulary, not a mechanical Simplified conversion. Transcription language now follows the interface language instead of a fixed auto, which had caused Chinese to be misread as Japanese. Language preference loads before first render, removing the startup flash. 5. Membership (1.4) react-native-iap, one-time lifetime purchase: buy, restore, entitlement validation, 10-day trial countdown, full-screen paywall, and a Settings upsell card that hides once VIP is active. 6. Other Advanced settings: auto-save interval, chunk length, and speaker sensitivity are configurable; interrupted transcriptions resume from completed chunks. Imported m4a files are transcoded to WAV via a new AVFoundation module. Speakers & history: save named speakers as voiceprint profiles, switch speaker mid-recording; filter history by speaker, search with highlighted matches, export results. Privacy & compliance: all recognition runs on-device; no audio, transcripts, or embeddings are uploaded. Recordings stay in the sandbox, leaving only when the user taps share. Third-party model names removed from user-visible copy. UI: icon-based transcript actions; export filenames use the recording title; fixed titles not refreshing after edit and speaker labels starting at 0.

VoxSafe 1.3 → 1.4: Feature Summary Version 1.4.0 brings a major diarization rewrite (sherpa-onnx), configurable recording/transcription settings, and advanced UI. Summary: 59 files changed; +3,101/-3,419 lines. ## NEW FEATURES 1. Speaker diarization rewritten with sherpa-onnx Replaced the old MFCC/ECAPA-based native module with a new voxsafe-sherpa-diarization module (pyannote segmentation + CAM++ embeddings via extractEmbeddings). New segment-based diarization pipeline: analyzeDiarizationFromSegments + identifySpeakers maps cluster centroids to saved speaker profiles (voiceprint matching). Added build scripts (build-sherpa-onnx-ios.sh) and Python verification (verify_sherpa_diarization.py). 2. Configurable Advanced Settings (new “Advanced” section) Recording auto-save interval (15/30/60 s) — replaces hardcoded 30 s. Transcription chunk size (15/30/60 s) — replaces hardcoded 30 s. Speaker sensitivity threshold (high/medium/low) — replaces hardcoded 0.5. All persisted via SecureStore; loaded at app startup. 3. Settings UI overhaul ## MODIFICATIONS HomeScreen: switched to analyzeDiarizationFromSegments; voiceprint matching only when profiles exist; added processing indicator, onNavigateSettings prop; upsertTranscriptSegments replaces insertTranscript. DetailScreen: transcript now merges multiple segments into one when no diarization speaker labels exist (prevents split artifacts). TranscribeService: chunk size is now configurable; unified speaker-diarization entry point (applySpeakerDiarization); uses computeHardBoundaries; removed analyzeDiarizationFull and mergeDiarizationWithSegments imports. RecordingService: recording persistence hardening — copy-to-persistent with periodic save (configurable interval), WAV header repair (fixWavHeader/findWavDataOffset), crash/kill recovery, persistent URI tracking, transcript audio URI linkage. WhisperService: model size verification rejects truncated downloads (95% threshold). SettingsScreen: +270 lines — Advanced section (3 parameter pickers), green privacy card, membership card, soft-sell banner. RecordButton/RecordingRow: icon-first action layout; RecordingDatabase +25; TranscriptDatabase +44 (upsert support). i18n: +39 keys each in en/zh (paywall, trial, advanced settings, privacy, membership); App.tsx wrapped with EntitlementProvider; PaywallScreen rendered at root with lock mode; all settings loaded at startup. app.json: +6 config changes; .gitignore: +5 entries (model binaries/artifacts). ## DELETIONS Entire voxsafe-diarization native module (old MFCC/ECAPA: 1,072-line Swift, 84 MB ONNX model, mlpackage, sb_cache). Entire voxsafe-speech-stream module (521-line Swift streaming speech recognition). SpeakerEmbeddingService and StreamingTranscribeService (superseded by sherpa-onnx pipeline); all streaming-transcription code paths removed — app now uses chunked Whisper transcription exclusively. Fixed a cross-repo typo (docslens → doclens) in FamilyLinkService and FeedbackService. Net result: 1.4 is a cleaner, monetized app with on-device voiceprint speaker identification, resilient recording persistence, and user-tunable transcription parameters.

1. Speaker Diarization Added speaker diarization powered by CoreML ECAPA-TDNN voice embedding model, supporting 2-6 speaker conversations. Uses K-Means clustering for accurate separation. Diarization runs in the background - transcripts save instantly, speaker labels update automatically when complete. Supports voice print enrollment and cross-recording identification. Per-segment speaker reassignment via person icon on each transcript segment. Settings includes max speaker selector (2-6) and voice print management. 2. Background Recording Recording continues when switching to background or lock screen. Auto-pauses during phone calls or other audio, auto-resumes when interruption ends. One-time notice on first use. Pre-recording model download check. 3. Streaming Transcription Whisper transcribes in 30-second slices during recording, merging full transcript on stop. Removed Apple Speech engine entirely, keeping only Whisper. 4. History Search & Filter Multi-dimensional search: keyword, speaker label, and time range (this week / this month / custom date). Keyword highlights in results. Filtered results support SRT/CSV export with _filtered suffix. 5. Family Link (JianBox Integration) Send transcripts and audio to JianBox via single merged URL (text + audio in one openURL). Archived recordings show gray badge. Receive URLs from CalMerge (voxsafe://new), JianBox (voxsafe://link), and DocLens (voxsafe://import). App Group shared files indexed by JianBox global search. Graceful degradation when App Group unavailable (free developer accounts). Concurrent file locking for safe dual-write. 6. Audio Interruption & Auto-Resume VoxSafeAudioInterrupt module handles audio session interruptions, remote control commands, and Now Playing info. Auto-resume recording after interruption ends. 7. Playback Improvements Waveform visualizes playback progress: played=blue, unplayed=dark gray, current=white with playhead. Tap waveform to seek. Three-state player: pause / resume / replay. Transcription slice checkpoint/resume for interrupted transcriptions. 8. Feedback Feature Settings page feedback entry with rich text editor (bold, italic, screenshot insertion), optional email and device info, one-tap submit. 9. UI Polish Transcription action buttons changed to vertical icon layout in title row. Recording title defaults to locale-formatted date (Chinese: 2026/08/02 14:30:45, English: 08/02/2026, 02:30:45 PM). Imported recordings use filename as title. Export filenames use recording title instead of duration/ID. Title edits refresh UI immediately. Share audio uses title as filename. 10. Stability Fixes Fixed crash from Whisper and CoreML GPU conflict (CoreML now uses CPU + Neural Engine). Fixed broken recording paths after app reinstall. Fixed speaker labels starting from 0 (now consistently Speaker 1+, with data migration). Improved diarization performance by eliminating redundant VAD passes. Auto-export to shared directory after transcription completes.

What's New in Version 1.2 NEW FEATURES • Automatic Speaker Diarization — VoxSafe now identifies and labels different speakers in a conversation automatically, using on-device pitch analysis. No cloud processing required. Configure the maximum number of speakers (2–4) in Settings, or switch to manual speaker selection during recording. • Editable Recording Titles — Name your recordings when you save them. A title prompt appears right after you stop recording, with a smart default. Edit titles anytime from the detail screen. • SRT & CSV Export — Export transcripts as SRT subtitle files or CSV spreadsheets. Perfect for video subtitling, legal documentation, or importing into other tools. • Recording Import — Import audio files from the Files app or any document provider. VoxSafe transcribes imported audio just like a live recording. • Real-Time Markers — Drop timestamped markers during recording to flag key moments. Each marker supports a short note for quick reference later. • Full-Text Search — Search across all transcripts from the history screen. Find any conversation in seconds. • In-App Feedback — Send bug reports and suggestions directly from Settings. IMPROVEMENTS • Lock-screen pause/resume control with Now Playing info • Background transcription keeps running when you leave the app • Transcription status tracking (pending, transcribing, completed, interrupted, failed) • Automatic recovery of incomplete recordings after force-quit • Bottom tab bar for quick navigation between Home, History, and Settings • Cleaner, more focused recording screen PRIVACY All features run entirely on-device. Speaker diarization uses local pitch analysis — no audio is uploaded. Whisper transcription is fully offline. Your data stays yours.

VoxSafe 1.1 New in this update: • Pause & Resume Recording — Tap once to pause, double-tap to finish. Recording auto-pauses when you take a call or switch apps, then picks right back up. • Face ID Lock — Protect your recordings with Face ID or a PIN. VoxSafe locks automatically whenever it goes to the background. • Smarter Transcription — Whisper transcription now includes precise timestamps and speaker segments for cleaner, more readable transcripts. • Offline Speech Recognition — New toggle to run Apple speech recognition fully on-device. Your audio never leaves your iPhone. • Silence Detection — Recordings with no speech are now saved and clearly labeled instead of being reported as failed. • Faster, more reliable audio engine and various bug fixes. Everything stays on your device. Always.