Each dictation previously spawned whisper-cli with -m, reloading the full
1.5-1.6GB model (+ ~10s Metal init) every single time. This adds a
WhisperServer manager that keeps a whisper-server child process alive with
the model preloaded, so each utterance is a fast local HTTP POST to
/inference (~0.5s) instead of a multi-second cold start.
- Bundles whisper-server alongside whisper-cli (bundle-whisper.sh).
- Server (re)starts on launch and on model change; stops on terminate;
picks a free loopback port; -nt (plain text) + -sns (fewer hallucinations).
- transcribe() prefers the warm server and falls back to a per-call
whisper-cli spawn if the server is unavailable.
- Warm-up now starts BEFORE the blocking accessibility prompt so the model
load overlaps with the user granting permissions.
- Info.plist: NSAllowsLocalNetworking for the loopback HTTP call.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Features:
- whisper-cli and dependencies bundled in app (no brew required for users)
- Built-in model downloader with 5 model options (Tiny to Large)
- Download progress indicator (e.g., "45% Downloading Small")
- Model selection dropdown in Settings
- First-run wizard for initial model download
Security fixes:
- Input validation for language and model path
- Private temp directory for audio files
- Auto-cleanup of audio files after transcription
- Auto-detect whisper-cli path (ARM/Intel)
Bug fixes:
- Fixed Settings window crash (NSWindowController)
- Fixed model selection with tag-based indexing
Build:
- bundle-whisper.sh script for packaging whisper-cli
- Updated Makefile with Frameworks bundling
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
