Simple Meeting Scribe

Fork notice. This is a personal fork of
czlonkowski/simple-meeting-scribe
by Romuald Członkowski, used under the MIT License (unchanged — see
LICENSE). All credit for the original app is his.

What this fork adds/changes: Ukrainian support (replaces Polish);
one-tap summary export to Telegram / Obsidian / email; persistent voice
enrollment that auto-labels known speakers across meetings; a
search_transcripts MCP tool; optional live on-screen subtitles via the
Gemini Live API; calendar integration for attendee-based speaker names;
menu-bar-only operation; and assorted recording/stability fixes. The
sections below are largely the original author's and still describe the
shared core.

A personal, 100% local meeting transcriber for macOS. I built it for myself.
I'm putting the source out there because other people asked — not because
I'm trying to ship a product.

There is no website, no installer, no support, no roadmap. If something
breaks, you fix it. If you want a feature, you add it. The project is MIT
licensed; fork it, strip it, reshape it — it's yours.

What it does

• Detects when a Google Meet / Zoom / Teams / Whereby call opens in Arc,
  Safari, or Chrome and offers to start recording.
• Records your microphone and the system audio of the meeting as two
  separate tracks (so the transcript can label "You" vs "Remote").
• Transcribes both tracks with WhisperKit
  (Whisper Large v3 / v3 Turbo, CoreML / ANE).
• Diarizes the remote track with FluidAudio
  (Pyannote on CoreML) so multiple remote speakers get distinct labels.
• Saves .md + .json transcripts to ~/Documents/MeetingTranscripts/.
• Optional on-device LLM summarization + action items + auto-titles via
  MLX — Bielik for Polish,
  Qwen3.5 for English.
• Drag any .mp4 / .m4a / .mov / .wav / .mp3 onto the window to
  transcribe an existing recording.

What this fork adds

On top of the original, this fork adds:

• Ukrainian support — replaces Polish. Transcription runs Whisper
  large-v3 for Ukrainian (turbo for English) by default; a per-term glossary
  • word-replacement pass cleans up domain jargon Whisper can't infer.
• Summary export — send a generated summary straight to a Telegram
  chat (Bot API), an Obsidian vault (as a dated markdown note), or a
  pre-filled email draft. Configure the destinations in Settings; nothing
  is sent unless you opt in.
• Voice enrollment — remember a speaker's voice once ("Andrii") and future
  meetings auto-label the same voice instead of "Remote 1/2". Uses FluidAudio's
  256-dim speaker embeddings + cosine matching, all on-device.
• Transcript search — a search_transcripts MCP tool for full-text search
  across every stored transcript, so an MCP client (e.g. Claude Code) can
  answer "what did we decide about X?" across all your meetings.
• Live subtitles — optional on-screen captions during a call via the Gemini
  Live API. Off by default; the app stays fully local otherwise.
• Calendar integration — pulls attendee names from the live calendar event
  as candidate speaker names.
• Menu-bar-only operation, plus recording/stability fixes (system-audio tap
  recovery, crash-safe recovery of interrupted recordings, arm64 build fixes).

What it does not do

• No cloud. No telemetry. No account. No network calls except the first-run
  downloads of Whisper / diarizer / LLM weights from Hugging Face.
• No auto-update. No App Store. No notarized binary (build it yourself).
• No tests. No CI. No documented API. No backwards-compatibility promise.
• Not localised beyond English + Polish (the two languages I need).

Everything happens on your machine. If the network is off, models already
downloaded keep working.

Requirements

• Apple Silicon Mac (M1 or newer; M3+ strongly recommended for the 11B
  Polish model). Intel is not supported — MLX and WhisperKit both assume
  Apple Silicon.
• macOS 26 Tahoe. The UI uses Liquid Glass, @Observable, and other
  macOS 26 APIs. Older macOS will not build.
• Xcode 16+ and the Xcode Metal Toolchain (Xcode will prompt on first build,
  or you can pre-install with xcodebuild -downloadComponent MetalToolchain).
• xcodegen — brew install xcodegen.
• ~15 GB free disk space if you want to cache all the optional models.
• 32 GB RAM recommended for the larger LLMs (Qwen3.5-9B, Bielik-11B). 16 GB
  works for Qwen3.5-4B and the smaller Bielik variant.

Build from source

git clone https://github.com/czlonkowski/simple-meeting-scribe
cd simple-meeting-scribe
xcodegen generate
xcodebuild -project MeetingTranscriber.xcodeproj \
           -scheme MeetingTranscriber \
           -configuration Debug \
           -destination 'platform=macOS' \
           -skipMacroValidation \
           build

Run the app from ~/Library/Developer/Xcode/DerivedData/.../Debug/MeetingTranscriber.app,
or copy it into /Applications with sudo cp -R ….

…or let an AI coding agent install it for you

If you already use Claude Code, Codex, or a
similar coding agent, paste this prompt into it and let it do the work. You'll
still need to approve Xcode / Homebrew / sudo prompts as they come up.

Please install Simple Meeting Scribe on this Mac. It's a SwiftUI app
at https://github.com/czlonkowski/simple-meeting-scribe. Do this end
to end:

1. Verify prerequisites: Apple Silicon, macOS 26 or newer, Xcode 16+
   installed. If Xcode Command Line Tools aren't installed, run
   `xcode-select --install` and wait for it to finish.
2. Install xcodegen if missing: `brew install xcodegen` (install
   Homebrew first with the official script if it's not there).
3. Pre-download the Metal Toolchain so the build doesn't stall on it:
   `xcodebuild -downloadComponent MetalToolchain`
4. Clone the repo into ~/Developer (create the directory if needed)
   and `cd` into it:
   `git clone https://github.com/czlonkowski/simple-meeting-scribe
   ~/Developer/simple-meeting-scribe && cd ~/Developer/simple-meeting-scribe`
5. Generate the Xcode project: `xcodegen generate`
6. Build Release:
   `xcodebuild -project MeetingTranscriber.xcodeproj
   -scheme MeetingTranscriber -configuration Release
   -destination 'platform=macOS' -skipMacroValidation
   -derivedDataPath build build`
7. Install to /Applications (this needs sudo — ask me to run it if you
   can't):
   `sudo rm -rf /Applications/MeetingTranscriber.app &&
    sudo cp -R build/Build/Products/Release/MeetingTranscriber.app
    /Applications/ &&
    sudo /System/Library/Frameworks/CoreServices.framework/Frameworks/LaunchServices.framework/Support/lsregister
    -f /Applications/MeetingTranscriber.app`
8. Open the app from /Applications once so macOS can register it, then
   tell me:
   - to grant Microphone + Screen Recording when prompted,
   - to approve Automation access for Arc / Safari / Chrome on first
     meeting detection,
   - to open Settings → Summary → Model Library and download a model
     before the first summarize (Qwen3.5-4B 8-bit for English,
     Bielik-11B v3 for Polish are the defaults).

If any step fails, stop and show me the exact error — don't paper
over it. If a step asks for sudo, run it only once and with my
permission.

Why -skipMacroValidation? MLX Swift LM ships compiler-plugin macros
(#hubDownloader, #huggingFaceTokenizerLoader). Xcode prompts to trust
them on first use. The CLI flag skips that prompt. In Xcode GUI, click
"Trust & Enable" on first build instead.

Because the app is ad-hoc signed, macOS will treat each install path as a
separate TCC identity. If you move the .app from DerivedData to
/Applications, you'll be asked to re-grant Microphone + Screen Recording

• Automation (for browser AppleScript). This is macOS behaviour, not mine.

First run

1. Launch. The dock icon is generated by scripts/generate_app_icon.swift —
   regenerate any time with swift scripts/generate_app_icon.swift.
2. Start a recording once; macOS will pop prompts for Microphone and
   Screen Recording. Approve both.
3. Open a meeting URL in Arc/Safari/Chrome. macOS will pop an Automation
   prompt for each browser the first time the app queries it.
4. Settings → Summary → Model Library → pick a model, click Download.
   Models live under ~/Documents/huggingface/models/.

Files on disk

• ~/Documents/MeetingTranscripts/ — transcripts (.md + .json).
• ~/Documents/MeetingTranscripts/recordings/ — the paired .voice.wav
  and .system.wav files for each session.
• ~/Documents/huggingface/models/ — downloaded LLM weights.
• Whisper models live under WhisperKit's own cache (first download shows
  progress inside the app).

Deleting a transcript from the sidebar (right-click → Delete) removes all
of the above for that session.

Repo layout

MeetingTranscriber/
├── App/             # SwiftUI @main + AppState + menu-bar extra
├── Capture/         # AudioRecorder, SystemAudioCapture (SCStream), MeetingDetector
├── Transcribe/      # WhisperEngine, DiarizationEngine, TranscriptionPipeline
├── Summarize/       # MLX wrapper, per-language prompts, model enum
├── Storage/         # TranscriptStore, DictionaryStore, SummaryStore
├── UI/              # SwiftUI views
└── Resources/       # Info.plist, entitlements, meeting-patterns.json

project.yml drives everything — edit the YAML, run xcodegen generate,
never touch the .xcodeproj by hand.

Contributions

I'm not taking pull requests. If you want to add something, fork it.
Issues are welcome as a place to discuss, but I make no promise to
respond.

License

MIT — see LICENSE.