Real-time voice transcription for macOS — 100% local, zero cloud.
IciTranscript captures your microphone and system audio simultaneously, transcribes everything in real time using on-device MLX models, and optionally generates an AI summary — all without sending a single byte to the cloud.
- Real-time transcription — see your words appear as you speak, powered by voxmlx and Apple Silicon MLX
- Dual audio capture — microphone + system audio (meetings, videos, calls) captured and mixed simultaneously
- Session history — every transcription is automatically saved and instantly visible in the sidebar; click any session to review it
- Speaker diarization — after a session, remote participants (system audio) are separated and labelled, up to 8 speakers, with NVIDIA Nemotron 3 Diarization running on MLX (nemotron-diar-mlx) and word-level timestamps from Parakeet; the model (~400 MB) is downloaded on first use
- AI summaries — get a concise summary of any session using a local Mistral model; summaries are saved and viewable directly in the session detail panel
- Zero setup for AI — Ollama runtime (~106 MB) and the Mistral model (~4 GB) are downloaded automatically on first launch, with a detailed progress screen; no manual steps required
- Export — copy transcript to clipboard or export as Markdown to
~/Downloads/ - Delete on hover — remove any session directly from the sidebar with a single click
- 100% local & private — your audio never leaves your Mac; no accounts, no subscriptions, no telemetry
- Keyboard shortcuts — start/stop recording from anywhere on your system
- Clean, native macOS UI — built with Flutter, feels at home on macOS 13+
| Component | Minimum | Recommended |
|---|---|---|
| macOS | 13 Ventura | 14 Sonoma or later |
| Chip | Apple M1 | Apple M2 / M3 |
| RAM | 8 GB | 16 GB |
| Disk | 5 GB free | 10 GB free |
Why Apple Silicon? The transcription engine uses MLX, Apple's machine learning framework optimised for the M-series neural engine. Intel Macs are not supported.
Enable the summary toggle before stopping a session. On first launch, the app automatically:
- Shows a dedicated setup screen with step-by-step progress
- Downloads the Ollama runtime (~106 MB)
- Starts the Ollama server
- Pulls the Mistral model (~4 GB) — one-time download, progress shown per step
After that, summaries are generated entirely on your Mac with no internet connection required. Summaries are saved per session and viewable by clicking the session in the sidebar. Everything is stored in ~/Library/Application Support/IciTranscript/.
- Download the DMG from GitHub Releases
- Drag IciTranscript into your Applications folder
- Right-click › Open on first launch (Gatekeeper warning — the app is ad-hoc signed)
- Grant Microphone and Screen Recording permissions when prompted
uv (the Python package manager used to run the transcription server) is installed automatically on first launch.
- Flutter SDK 3.x
- Xcode 15+
uv— astral.sh/uv
# Install Flutter dependencies
cd apps/app_mobile && flutter pub get
# Run in debug mode
flutter run -d macos
# Run tests
flutter test
# Analyse
flutter analyze
# Build release
make build
# Create DMG
make dmgapps/app_mobile/ — Flutter macOS application
packages/
core_domain/ — Entities, use cases, repository interfaces
core_data/ — Repository implementations, Drift SQLite database
core_foundation/ — Logger, config, shared utilities
core_presentation/ — Shared widgets
IciTranscript follows a layered clean architecture:
- Presentation — Flutter widgets + Riverpod ViewModels (Freezed state)
- Application — LiveTranscriptionService orchestrating audio capture, WebSocket, and DB
- Domain — Pure Dart use cases and entities
- Data — Drift SQLite + platform channels (Swift) for audio + process management
The transcription server (voxmlx-serve) runs as a child process launched by the app via uvx. It listens on ws://localhost:8000 and streams transcript segments back in real time.
IciTranscript is designed with privacy as a core principle:
- All audio processing happens on your device
- No data is sent to any external server
- No analytics, no crash reporting, no telemetry
- Session data is stored locally in
~/Library/Application Support/IciTranscript/
MIT