Meeting Transcriber & Summarizer

This project was created to help me transcribe and summarize meetings, including both:

• Online meetings (by recording system audio)
• In-person or hybrid meetings (by switching to microphone input instead)

Everything here is built with free and open source software and designed to run entirely offline, including transcription and summarization. This is especially helpful for privacy-sensitive contexts like interviews or internal team discussions.

---

Features

• Record system audio (or mic)
• Transcribe using OpenAI Whisper
• Summarize using Mistral-7B Instruct via llama-cpp-python
• CPU-only — no GPU required
• Auto-saves recordings, transcripts, and summaries in session-specific folders
• Hackable, scriptable, fully offline

---

Tested Environment

• OS: Ubuntu 20.04.6 LTS
• Python: 3.11
• Virtual environment: created with venv
• Hardware: CPU-only (no GPU)

---

Setup Instructions

1. Install system dependencies

Before running the scripts, make sure you have the following system tools installed:

sudo apt install ffmpeg pulseaudio-utils

These are used to capture system audio (parec from PulseAudio), convert audio formats (ffmpeg), and list available devices (pactl).

Create the required folders if they don't already exist:

mkdir -p recordings transcripts summaries

2. Clone this repo

git clone https://github.com/yourusername/meeting-transcriber.git
cd meeting-transcriber

3. Create virtual environment

python3.11 -m venv audio_venv
source audio_venv/bin/activate

4. Install dependencies

pip install -r requirements.txt

5. Download models

• Whisper (automatic on first run)
• Mistral 7B Instruct (Q4_K_M)
  • Download from TheBloke/Mistral-7B-Instruct-GGUF
  • Place the .gguf model in the root directory and update the path in summariser.py

---

How to Run

Full Pipeline (One Command)

python run_meeting_pipeline.py

• Prompts you to name the meeting
• Records until you press Enter
• Make sure the source monitor is correct (edit record_system.py, update the parec -d device if needed)
• Use pactl list sources short to find the correct monitor
• Automatically transcribes and summarizes it
• Saves everything in: -./recordings/<name>_audio.wav -./transcripts/<name>_transcript.txt -./summaries/<name>_summary.txt

Output File Structure

recordings/
  meeting123_20250722_audio.wav

transcripts/
  meeting123_20250722_transcript.txt

summaries/
  meeting123_20250722_summary.txt

---

Future Work / Improvements

• Speaker diarisation (label different speakers)
  • Might require some level of user input
• Make it more usable for non-devs (e.g., GUI or CLI prompts)
• No hardcoding: auto-detect system source instead of manually editing record_system.py
• Maybe deploy it as a simple app, browser extension, or Electron GUI
• Possibly explore real-time meeting assistant extensions

---

Contributions

Pull requests welcome — this is a very rough but functional start. If you have ideas for better UX, diarisation, efficient streaming, real-time handling, etc., feel free to contribute.

---

License

This project is fully open source and licensed under the MIT License.

---

Built for people who prefer local control, privacy, and open tools.

Acknowledgements

• OpenAI Whisper for transcription
• Mistral-7B + llama-cpp-python
• Ubuntu + PulseAudio + parec/ffmpeg ecosystem
• Hugging Face community for incredible model access

---

Happy hacking!