MeetingBot

A fully local, zero-cloud meeting intelligence system for macOS.

Records Google Meet audio + your microphone, then automatically:

• Transcribes with Whisper large-v3 (Apple Silicon MPS accelerated)
• Diarizes speakers with pyannote 3.1
• Summarizes with a local LLM via Ollama (llama3.1:8b)
• Exports a 10-section dark-theme HTML report

Supports Hindi, English, and Hinglish code-switching. Zero data leaves your Mac.

---

Quick start

Prerequisites

• macOS 13+ (Ventura or Sonoma)
• Apple Silicon or Intel Mac with 16 GB RAM recommended
• Python 3.11+
• Homebrew

Install

git clone https://github.com/Ashutosh-code-arch/meeting-bot
cd meeting-bot
bash install.sh

Configure (one-time)

# 1. Edit .env — add your HuggingFace token
nano .env

# 2. Accept pyannote model terms (one-time browser step)
open https://hf.co/pyannote/speaker-diarization-3.1

# 3. Audio MIDI Setup — route Meet audio through BlackHole
open -a "Audio MIDI Setup"
# Press (+) → Create Multi-Output Device
# Check: BlackHole 2ch + Built-in Output
# In Google Meet: Settings → Audio → Output → Multi-Output Device

Run

# Terminal 1: start the tray app (or use launchd for auto-start)
make run

# Terminal 2: start the background worker
make worker

Then in the menu bar: click MeetingBot → Start Recording before your Meet call. Click Stop and Process when the meeting ends. A browser window opens automatically with the report.

---

Development

make test        # unit tests (no hardware needed)
make test-int    # integration tests (needs BlackHole + Ollama)
make lint        # ruff linting
make check       # verify Ollama + audio devices are ready
make ls          # list all recorded meetings
make logs        # tail worker logs live

---

Project structure

meetingbot/
├── meetingbot/       Python package (all pipeline code)
├── prompts/          LLM system prompt (edit to customise output)
├── templates/        HTML report Jinja2 template
├── tests/            pytest test suite
├── install.sh        One-time setup script
├── install_launchd.sh Auto-start on login setup
├── Makefile          Common dev commands
└── .env              Your secrets (HF_TOKEN etc.)

---

Configuration

All settings in .env:

Variable	Default	Description
HF_TOKEN	(required)	HuggingFace token for pyannote
WHISPER_MODEL	large-v3	Use medium if RAM limited
WHISPER_LANG	hi	hi handles Hinglish best
OLLAMA_MODEL	llama3.1:8b	Use 70b for better quality
SYSTEM_DEVICE	BlackHole	Virtual audio device name
MIC_DEVICE	MacBook	Your mic device name fragment
MEETINGBOT_DIR	~/meetings	Storage directory

---

Output

Each meeting produces:

• ~/meetings/reports/mtgXXXX_YYYYMMDD.html — dark-theme browser report
• ~/meetings/reports/mtgXXXX_YYYYMMDD.md — raw Markdown

Report sections: Transcript · Summary · Speaker-wise Summary · Tasks Table · Feature Plan · Day-wise Plan · Implementation Notes · Risks · References · Final Notes

---

Troubleshooting

Problem	Fix
BlackHole not capturing Meet	Set Meet audio output to Multi-Output Device in Meet settings
HF_TOKEN error	Accept model terms at hf.co/pyannote/speaker-diarization-3.1
Ollama not responding	Run: brew services start ollama
MPS not available	Check: python -c "import torch; print(torch.backends.mps.is_available())"
Mic permission denied	System Settings → Privacy → Microphone → allow Terminal
Out of memory	Switch to WHISPER_MODEL=medium in .env
Worker not processing	Check: tail -f ~/meetings/worker.log

---

Roadmap

• Phase 1 (this repo): Local Mac app — one user, one machine
• Phase 2: Self-hosted Docker service — multi-user, REST API
• Phase 3: Desktop client + encrypted cloud sync