EXCOM 76 2025

This repo host the scripts used to perform the transcript and analysis of the 76th Session of UNHCR Executive Committee.

At its annual session, the Executive Committee (ExCom) reviews the work of the Standing Committee of the preceding year and establishes its programme of work for the next 12 months.

Each session of the debate is available within the UN Web TV:

Day 1 - 6 Oct 2025

1. AM
2. PM

Day 2 - 7 Oct 2025

3. AM
4. PM

Day 3 - 8 Oct 2025

5. AM
6. PM

Day 4 - 9 Oct 2025

7. AM
8. PM

Day 5 - 10 Oct 2025

9. AM
10. PM

This about 30 hours of discussions... A good use case for using AI!

Step 0 - Environment

Set up a python virtual environment and install requirements

uv venv
uv pip install -r requirements.txt

Then set up your env file with API key and setting following the .env.template

Note that this project was implemented on a linux env with a gpu with cuda for shorter processing.

Step 1 - Transcription with AI

1. Each video (english version) was locally recorded with ffmpeg -f pulse -i alsa_output.pci-0000_00_1f.3-platform-skl_hda_dsp_generic.HiFi__hw_sofhdadsp__sink.monitor audio/session1.wav.

2. A first script uv run python3 script1_transcribe.py --mode single --input audio/session1.wav or uv run python3 script1_transcribe.py --mode batch --input audio/ --output transcript/batch_results_d.txt --model-size small --device cuda to transcribe the audio recording with whisper model.

3. This was manually re-associated with the official list of speakers. The output is transcript/excom_transcript.json which is then converted into a PDF document with uv run python3 script2_trans_to_pdf.py

Step 2 - Analysis with AI

The script uv run python3 script2_analyze.py use the transcript to generate a structured report out of all the transcripts. The output are both a report.md and report.docx

Step 3 - Disseminate with AI

A dedicated script uv run python3 app.py was built to publish the report - this app not only allows to read the report but provide an additional bot functionnality that allows to know more about each part of the report.

A precompiled list of follow up questions are offered for each elements of the report - and results are generated using Retrieval Augmented Generation based on the entire documentation - PDF files in a folder call docs that was provided for the EXCOM

Using the app, Users get then the ability to dig further on any specific points that was detected within the debate with answers grounded on all the documentation provided for the meeting.