Meeting Scribe

Transcribes meeting audio and generates structured minutes with action items. Uses OpenAI Whisper for transcription and your choice of model for the minutes:

• GPT-4.1 Mini (OpenAI, closed-source) — fast, no GPU required
• Llama 3.1 8B (Meta, open-source) — runs locally, requires a CUDA GPU

Setup

1. Install dependencies

uv sync

For the open-source Llama path, also install the llm extras:

uv sync --extra llm

2. Configure environment

cp .env.example .env

Then edit .env:

OPENAI_API_KEY=sk-...       # required for Whisper transcription and GPT minutes
HF_TOKEN=hf_...             # required only when using the Llama model

To use Llama 3.1 8B you also need to accept the license on HuggingFace.

Note: You do not need to download the model manually. On the first run with Llama selected, the weights (~16 GB) are downloaded automatically to your HuggingFace cache (~/.cache/huggingface/). Subsequent runs use the cached copy. A CUDA-capable NVIDIA GPU is required.

3. Run

uv run meeting-scribe

Then open the local URL printed to the terminal (default: http://127.0.0.1:7860).

Usage

1. Upload an MP3 or WAV recording of your meeting
2. Select a model (GPT-4.1 Mini or Llama 3.1 8B)
3. Click Generate Minutes

The app returns:

• Full transcript of the audio
• Markdown meeting minutes with summary, key discussion points, takeaways, and action items with owners

Project structure

src/meeting_scribe/
├── transcribe.py   # Whisper transcription via OpenAI API
├── minutes.py      # Minutes generation (GPT and Llama backends)
└── app.py          # Gradio UI and entry point

Sample audio

The original course used a Denver City Council meeting extract. You can download it here or find the full dataset on HuggingFace.