Mila

A native macOS app that records, dictates, and transcribes locally on your Mac — no audio leaves the device. Hebrew transcription is powered by the ivrit.ai large-v3 finetune of Whisper, English by OpenAI's large-v3-turbo, both running through whisper.cpp (GPU via Metal).

Features

• Record from the microphone at any time.
• Record system audio from a single app (Zoom, Google Meet in Chrome, Slack, etc.) via ScreenCaptureKit — no virtual audio device required.
• Record meetings = mic + system audio mixed into one mono 16 kHz WAV file.
• Dictate in two languages with separate global hotkeys:
  • ⌘2 — English dictation (OpenAI large-v3-turbo)
  • ⌘3 — Hebrew dictation (ivrit.ai large-v3) Both hotkeys are user-configurable in Settings → Hotkeys.
• Transcribe locally: ivrit.ai large-v3 (Hebrew) or OpenAI large-v3-turbo (English / multilingual). Audio never leaves your Mac.
• Per-segment timestamps, click to seek, copy/share transcript, RTL rendering for Hebrew.

System requirements

To run Mila:

• macOS 14 (Sonoma) or later.
• Apple Silicon (M-series) strongly recommended. Transcription runs on the Metal GPU (whisper.cpp) and speaker diarization on MPS/CPU (pyannote). Intel Macs fall back to CPU and are much slower.
• Disk: ~4.6 GB for the two default Whisper models, downloaded on first launch (ivrit.ai large-v3 ~3.0 GB + OpenAI large-v3-turbo ~1.6 GB). Add ~1 GB more if you enable speaker diarization (bundled Python runtime plus a torch download on first enable).
• Memory (approximate): 16 GB unified memory recommended. 8 GB is workable for plain transcription but tight with diarization. Rough working set: Hebrew large-v3 ~3.5–4 GB, English large-v3-turbo ~1.5–2 GB, speaker diarization adds ~1–2 GB. (RAM figures are approximate guidance, not a hard spec.)
• Live (real-time) mode — running transcription and diarization concurrently in real time is the heaviest path. It's automatically disabled on MacBook Air–class chips (they can't keep up in real time) and recommended on M-series Pro/Max, or M2 and newer. Recording plus after-the-fact transcription still works on Air.

To build Mila:

• Xcode 15.3 or newer (Command Line Tools alone are not enough).
• Homebrew (used to install xcodegen).

Build & run

From the project root:

make bootstrap   # one time: installs xcodegen via brew if missing
make project     # generates Mila.xcodeproj from project.yml
make open        # opens Xcode

Or do it all from the CLI:

make run

The first build will resolve Packages/WhisperBinary, which downloads the whisper.xcframework (~65 MB) from the official ggml-org/whisper.cpp v1.8.4 release. The checksum is pinned in Packages/WhisperBinary/Package.swift.

Models

On first launch the app downloads both default models in the background:

• ivrit-ai/whisper-large-v3-ggml — Hebrew (~3.0 GB). Empirically more accurate on Hebrew speech than the smaller large-v3-turbo finetune, which is why we ship the larger one despite the size and ~2× inference cost.
• openai whisper-large-v3-turbo (the ggerganov/whisper.cpp build) — English / multilingual (~1.6 GB).

You can monitor progress in the in-app banner or pre-download from the CLI:

make models

Models live at:

~/Library/Application Support/Mila/Models/
    ivrit-ai-whisper-large-v3.bin
    openai-whisper-large-v3-turbo.bin

The app picks them up from that directory automatically.

Remote transcription (optional)

By default Mila transcribes entirely on-device — audio never leaves your Mac. If you'd rather offload transcription (e.g. on a low-powered Mac, or to use a Hebrew-tuned model that's too large to run locally), Settings → Models lets you switch the backend to any OpenAI-compatible /v1/audio/transcriptions endpoint: OpenAI's own Whisper API, or a self-hosted server. The API key is stored in your Keychain, and the UI flags that audio leaves the device while this backend is active.

To run ivrit.ai's Hebrew models (or any other Whisper model) behind this API on your own machine, see docs/REMOTE_TRANSCRIPTION_SERVER.md.

Required permissions

On first use macOS will prompt for the following — all are required for the relevant feature to work:

Feature	System Settings panel
Microphone recording	Privacy & Security → Microphone
System / app audio	Privacy & Security → Screen Recording
Dictation paste-at-cursor	Privacy & Security → Accessibility

How meeting capture works

ScreenCaptureKit allows audio capture per-application. When you pick Zoom in the New recording sheet, only Zoom's audio output is captured (not your own mic; not other apps). The mic is captured separately via AVAudioEngine and the two streams are downmixed (averaged) into a single mono 16 kHz WAV file that's fed into Whisper. SCK requires a video stream to function, so a 2×2 placeholder is configured and discarded.

Project layout

Mila/
├── App/MilaApp.swift          # SwiftUI @main entry + AppDelegate
├── Audio/
│   ├── AudioUtilities.swift            # Conversion to 16k mono Float32
│   ├── MicrophoneRecorder.swift        # AVAudioEngine input tap
│   ├── SystemAudioRecorder.swift       # ScreenCaptureKit per-app audio
│   └── RecordingSession.swift          # Mix mic + system → WAV
├── Transcription/
│   ├── ModelManager.swift              # Download / select ggml models
│   ├── WhisperEngine.swift             # whisper.cpp C bridge
│   └── TranscriptionService.swift      # Orchestrates jobs
├── Dictation/
│   ├── HotkeyManager.swift             # Carbon global hotkeys (multi-binding)
│   └── DictationController.swift       # Mic → Whisper → paste
├── Views/                              # SwiftUI screens incl. Settings
├── Models/Recording*.swift             # Persisted recording metadata
└── Resources/
    ├── Info.plist
    └── Mila.entitlements
Packages/WhisperBinary/                 # SPM wrapper for whisper.cpp xcframework
scripts/make-dmg.sh                     # Builds a release DMG for distribution

Building a release DMG

# 1. Bump MARKETING_VERSION + CURRENT_PROJECT_VERSION in project.yml
# 2. make dmg                       # builds an ad-hoc-signed Mila-<version>.dmg

make dmg produces an ad-hoc-signed DMG suitable for local testing and self-distribution. macOS Gatekeeper will show the standard prompt on first launch (right-click → Open). To ship a notarized build, sign the app with your own Apple Developer ID and run it through Apple's notarytool — see the code signing notes below.

Code signing notes

project.yml sets CODE_SIGN_IDENTITY: "-" (ad-hoc) and disables sandbox so the dev build can use ScreenCaptureKit without provisioning. For distribution via the App Store you'd:

• Re-enable com.apple.security.app-sandbox
• Add the appropriate entitlements (com.apple.security.device.screen-recording, com.apple.security.network.client, etc.)
• Provide your team ID in DEVELOPMENT_TEAM.

Changelog

Newest first. Dates are release dates.

• Unreleased — Added an "Automatically summarize recordings" toggle (Settings → LLM, on by default) so Mila can run transcript-only: turn it off and no AI Overview is generated after a recording, on launch backfill, or on re-transcription. Added a configurable CLI timeout (default 300s, Settings → LLM) that applies to title generation, auto-summary, and the Send-action button; improved the timeout error message to name the elapsed seconds and point to the setting.
• 1.8.5 (2026-06-09) — Independent mic / app-audio capture toggles, AAC (.m4a) recordings, and a configurable recording-storage cap.
• 1.8.4 (2026-06-09) — Fixed speaker diarization silently breaking in notarized builds (the bundled Python couldn't load torch's dylibs).
• 1.8.3 (2026-06-08) — Mic-capture diagnostics, automatic language detection, and a post-recording sheet fix.
• 1.8.2 (2026-06-04) — Live transcription of app audio, screen-lock guard, recording-screen CPU fix, diarization improvements.
• 1.8.1 (2026-06-03) — Reverted automatic silence-dropping for app recordings.
• 1.8.0 (2026-06-01) — Five post-1.7 fixes: AGC-aware VAD, instant stop dialog, background recording mode, audio-context tuning.
• 1.7.0 (2026-05-28) — Automatic gain control, summary backfill, post-recording popup, and more.
• 1.6.5 (2026-05-28) — Removed a hardcoded email address from the default action prompt.
• 1.6.4 (2026-05-27) — Live speaker labels and instant save.
• 1.6.3 (2026-05-27) — Voice-activity-detection tuning for real conversations.
• 1.6.2 (2026-05-27) — VAD-driven live transcription, Claude Sonnet 4.6, rename sheet always shown.
• 1.6.1 (2026-05-27) — Live AI chunk size 5s → 30s to stop word-cutting.
• 1.6.0 (2026-05-26) — Live AI mode plus LLM/UI end-to-end test infrastructure.
• 1.5.0 (2026-05-25) — Speaker diarization on by default; added license and attributions.
• 1.4.3 (2026-05-23) — Build/CI maintenance (macOS 26 "Tahoe" runners).
• 1.4.2 (2026-05-23) — Restored the floating sidebar card and froze its material.
• 1.4.1 (2026-05-23) — In-place-update bundle relocator and UI polish.
• 1.4.0 (2026-05-23) — Stabilized the app bundle identifier.
• 1.3.11 (2026-05-20) — Speaker labels included in copied transcripts and LLM prompts.
• 1.3.10 (2026-05-19) — Diarization availability re-checked at launch to match what's on disk.
• 1.3.9 (2026-05-18) — Diarization self-heal: nuclear-repair fallback and matplotlib pre-install.
• 1.3.8 (2026-05-18) — Diarization self-heal: recover from missing Python modules; install numpy<2.
• 1.3.7 (2026-05-17) — Bundle torch dependencies; self-heal missing transitive deps.
• 1.3.6 (2026-05-17) — Release-packaging fix.
• 1.3.5 (2026-05-17) — Bundled diarization runtime (Python 3.11 + pyannote.audio) inside the app; manual rename and flat folders for organizing transcriptions.
• 1.3.1 (2026-05-13) — Maintenance release (version bump for auto-update).
• 1.3.0 (2026-05-13) — Initial speaker diarization and transcript (SRT) export.
• 1.2.8 (2026-05-13) — Security: pin and verify the SHA-256 of Whisper models downloaded from Hugging Face.
• 1.2.7 (2026-05-12) — Pick a specific input device; shell out to a local Claude/Cursor CLI to auto-name recordings.
• 1.2.6 (2026-05-11) — Fixed a Sparkle auto-update loop.
• 1.2.4 / 1.2.5 (2026-05-11) — Internal auto-update pipeline verification (no user-facing changes).
• 1.2.3 (2026-05-11) — Hardened the release DMG mount-path parsing.
• 1.2.2 (2026-05-10) — Stable code signing so in-place updates keep macOS permission grants.
• 1.2.1 (2026-05-10) — Fixed dictation pasting into the wrong window and a main-thread freeze with wireless mics.
• 1.2.0 (2026-05-09) — Switched the Hebrew default to the ivrit.ai large-v3 model (~3 GB).
• 1.1.0 (2026-05-09) — New app icon, per-language voice memos, hotkeys card, hide-recents toggle.
• 1.0.0 (2026-05-09) — Initial release.

License

Mila is licensed under the Apache License 2.0. See NOTICE for required attribution.

Originally developed by Uri Harduf at Island Technology, Inc.

Credits

• ivrit.ai for the Hebrew Whisper finetunes.
• whisper.cpp (MIT) for the local inference.
• OpenAI Whisper for the English model weights (re-packaged as ggml by ggerganov/whisper.cpp).
• Sparkle (MIT) for auto-updates.