About This Tool

EU-sovereign meeting notes โ€” built on the SURF AI Hub, no data leaves the EU.

๐ŸŽฏ Overview

The Meeting Notes Generator is a proof of concept application developed by the AI Expert Team at The Hague University of Applied Sciences (THUAS). It demonstrates how modern AI technologies can be combined to automate the tedious task of creating meeting notes from recordings.

Both speech-to-text and the language model that structures the notes run on the SURF AI Hub (WiLLMa), the sovereign AI inference platform operated by SURF for Dutch research and education institutions. The app is just an orchestrator โ€” no local GPU needed.

โš™๏ธ The Pipeline

๐Ÿ“

Upload

Video/Audio file

โ†’
๐Ÿ”Š

Extract Audio

FFmpeg processing

โ†’
๐ŸŽ™๏ธ

Transcribe

ASR Model

โ†’
๐Ÿค–

Process

LLM Analysis

โ†’
๐Ÿ“

Output

Structured Notes

๐ŸŽ™๏ธ Speech Recognition (via SURF AI Hub)

Whisper Dutch

  • Provider: SURF AI Hub (WiLLMa)
  • Model: openai/whisper-large-v2 (SURF AI Hub sequence id 7, display name "Whisper speech-to-text Dutch")
  • Languages: Dutch (primary), English
  • Hosted, OpenAI-compatible /audio/transcriptions endpoint
  • No local GPU required

The earlier on-device Parakeet (NVIDIA) and Voxtral-Mini (Mistral) backends are still in the code, commented out, as a fallback if SURF is unreachable.

๐Ÿค– LLM (via SURF AI Hub)

openai/gpt-oss-120b

SURF-hosted GPT-OSS 120B โ€” the LLM used for every meeting. Produces the structured summary, topics, decisions, action items, risks, and open questions.

Previously used local Ollama with Ministral models. Those code paths are commented out โ€” switching to SURF removed the local GPU requirement.

โœจ Features

  • ๐Ÿ“‹
    Structured Output

    Summary, Topics, Decisions, Action Items, Risks, Open Questions

  • โš™๏ธ
    Customizable Config

    Participants, glossary, meeting context for better accuracy

  • ๐Ÿ“
    Editable Prompts

    Customize how the LLM generates each section

  • ๐Ÿ“ฅ
    Multiple Formats

    Download as Markdown or DOCX

๐Ÿ› ๏ธ Technology Stack

Backend

FastAPI, Python

ASR

SURF AI Hub Whisper

LLM

SURF AI Hub ยท gpt-oss-120b

Audio

FFmpeg, SciPy

Frontend

HTML, CSS, JS

Export

python-docx

๐Ÿš€ Roadmap

โœ“
Core Pipeline

Upload โ†’ Transcribe โ†’ Generate Notes

โœ“
Multiple Models

Support for different ASR and LLM models

โณ
Speaker Diarization

Identify who said what

โ—‹
Desktop app (on-prem)

Package this into a desktop application so users can run the full pipeline locally โ€” fully on-prem, no cloud dependency. The tech stack will be updated for that.

๐Ÿ‘ฅ Credits

Built by Amey Vasulkar, working in the AI Expert Team at The Hague University of Applied Sciences (THUAS).

This is a proof of concept application for research and demonstration purposes. Inference is performed via the SURF AI Hub (WiLLMa), a sovereign AI platform operated by SURF for Dutch research and education.