MeetingMind — AI-Powered Meeting Intelligence System

Local-first meeting transcription, speaker diarization, and LLM-powered analysis. Runs entirely on your laptop — no cloud, no subscriptions.

Overview

MeetingMind is a multi-stage pipeline that captures audio, converts speech to text, assigns speaker identities, generates live summaries, and enables post-session querying via a retrieval-augmented generation (RAG) system.

Stack

Layer	Technology
Audio capture	sounddevice
Transcription	faster-whisper (CTranslate2)
Diarization	pyannote/speaker-diarization-3.1
LLM	Ollama (qwen2.5:3b by default)
Backend	FastAPI + WebSocket
Vector store	ChromaDB
Frontend	Plain HTML/JS (no build step)

Architecture

Audio Input → Audio Capture (chunking with overlap) → Transcription (Faster-Whisper) → Diarization (pyannote) → Speaker Resolution → Session Management → LLM Processing (live + post) → RAG Indexing → WebSocket Delivery to UI

Each stage is modular and can be tested independently.

Setup

1. Clone and create environment

cd meetingmind
python -m venv venv
source venv/bin/activate          # Windows: venv\\Scripts\\activate
pip install -r requirements.txt

2. Configure environment

cp .env.example .env
# Edit .env and add your HF_TOKEN

HF Token setup (one-time):

1. Create a free account at https://huggingface.co
2. Accept model terms at https://hf.co/pyannote/speaker-diarization-3.1
3. Accept model terms at https://hf.co/pyannote/segmentation-3.0
4. Generate a token at https://hf.co/settings/tokens
5. Paste it into .env as HF_TOKEN=hf_...

3. Pull Ollama models

ollama pull qwen2.5:3b          # live summaries
ollama pull nomic-embed-text    # RAG embeddings (post-session Q&A)

# Optional: larger model for post-session analysis
ollama pull llama3.1:8b

Running

cd backend
python main.py

Open http://localhost:8000 in your browser.

Usage

1. Enter known speaker names (comma-separated)
2. Press Start to begin recording
3. Speak; transcript appears in real time with speaker labels
4. Live summary refreshes periodically
5. Press Stop when the meeting ends
6. Press Analyse for post-session insights
7. Use the Q&A tab to query the transcript

Renaming speakers

Rename any anonymous speaker ("Speaker N") from the UI.

Testing individual components

cd backend
python audio_capture.py
python transcriber.py path/to/audio.wav
python diarizer.py path/to/audio.wav

Project Structure

meetingmind/
├── backend/
│   ├── main.py              # FastAPI app + WebSocket
│   ├── audio_capture.py     # Mic input → chunking
│   ├── transcriber.py       # faster-whisper wrapper
│   ├── diarizer.py          # pyannote diarization
│   ├── speaker_registry.py  # Cross-chunk speaker identity
│   ├── session_manager.py   # Session state + persistence
│   ├── llm_client.py        # Ollama client (live + post)
│   ├── rag.py               # RAG pipeline
│   └── config.py            # Configuration
├── frontend/
│   └── index.html           # Single-file UI
├── sessions/                # Stored transcripts
├── requirements.txt
└── .env

Performance Tips

Machine	Recommended Whisper model	Notes
MacBook (Intel)	base or tiny	CPU-only
GPU system	small or medium	Faster inference

Adjust in backend/config.py.

Limitations

• Near-live processing introduces latency
• Diarization consistency can drift across long sessions
• CPU inference may be slow on low-end machines
• Single active session at a time

Summary

MeetingMind converts conversations into structured, queryable data using speech recognition, speaker diarization, language models, and vector search. It is designed as a complete system blueprint with clear extension points for future development.