Meeting Recorder

Local-first macOS meeting recorder with on-device transcription and speaker recognition.

Features

• Record mic + system audio -- captures both sides of video calls via ScreenCaptureKit
• On-device transcription -- WhisperKit (CoreML) with model sizes from tiny (75 MB) to large-v3 (2.9 GB)
• Speaker diarization -- FluidAudio identifies who said what, with a learnable voice library that improves over time
• Source-aware speaker matching -- uses retained mic/system stems to reduce close-call speaker mixups
• Obsidian-compatible output -- structured markdown with YAML frontmatter and [[wikilinks]] for speaker pages
• Menu bar app -- lives in the menu bar with a global hotkey (Ctrl+Opt+R) for quick recording
• Crash recovery -- two-phase pipeline checkpoints expensive transcription work so crashes don't lose progress
• Auto-transcribe, auto-save -- configurable automation with retention policies (7-90 days)
• Fully local -- no cloud APIs, no internet required after initial model download

Requirements

• macOS 14+ (Sonoma)
• Apple Silicon (M1 or later)
• Xcode 15.3+ / Swift 5.10+ (for building from source)
• No Apple Developer account needed (ad-hoc code signing)

Build & Install

git clone https://github.com/tonton-golio/meeting-recorder.git
cd meeting-recorder/swift
./build.sh
open -a "Meeting Recorder"

The build script compiles via SPM, bundles a .app, ad-hoc signs it, and installs to /Applications. First build downloads dependencies and may take a few minutes.

On first launch, macOS will prompt for two permissions:

1. Microphone -- required for recording
2. Screen Recording -- needed to capture system audio (the remote side of video calls). Optional: the app falls back to mic-only if denied.

How it works

1. Record -- captures microphone audio (and optionally system audio) as a final 16kHz mono WAV, retaining mic/system stems while audio is kept
2. Transcribe -- WhisperKit runs Whisper locally via CoreML, producing timestamped text segments
3. Diarize -- FluidAudio clusters audio into speakers and extracts WeSpeaker embeddings (256-dim vectors)
4. Match -- each speaker embedding is compared against the People voice library using cosine similarity. Known voices are auto-labeled; unknown speakers are presented for identification.
5. Save -- the transcript is written as Obsidian-compatible markdown with YAML frontmatter

See docs/SPEC.md for the full product specification. See docs/SPEAKER_IDENTIFICATION_EXPERIMENTS.md for the source-aware matching experiment plan. See docs/SYSTEM_AUDIO_CAPTURE_EXPERIMENTS.md for system-audio capture diagnostics.

Data storage

All data lives locally under ~/.meeting-recorder/:

~/.meeting-recorder/
  recordings/
    recordings.json          # recording index
    *.wav                    # mixed audio files
    *.mic.wav                # microphone stems, while retained
    *.sys.wav                # system-audio stems, while retained
  people/
    people.json              # people index
    {uuid}/                  # per-person directory
      {sample-uuid}.caf     # voice samples
  meetings/
    *.md                     # saved markdown transcripts

Architecture

The app is a SwiftUI menu bar application with a central AppState coordinator that owns the recording, transcription, people-matching, and playback subsystems. See CLAUDE.md for a detailed architecture overview and component guide.

Acknowledgments

• WhisperKit by argmaxinc -- on-device speech-to-text via CoreML
• FluidAudio by FluidInference -- on-device speaker diarization and embeddings

License

MIT