Media Summary pipeline

This project aims at making a pipeline to takes detailed notes using chatGPT and Whisper. Whisper runs locally (in GPY if possible) to save cost.

Youtube URL --> Video --> Script --> Several block of texts --> several chatGPT 3.5 summary --> 1 concatenated big summary

Works for any media (audio, video) not only youtube.

1) Setup

Openai key :

In your bashrc file :

export OPENAI_KEY="sk-************************************"

Install the python env in conda

conda env create -f environment.yml

2) Download videos from youtube

Please note this pipeline can take as input any video / audio input, not only youtube video. Enter your url in the variable url_list of download_from_youtube.py

python download_from_youtube.py 

3) Call the pipeline

python main.py --input_dir_or_file /path/to/video_or_audio --final_output_dir /path_to_save_markdown 

Note that the chatGPT prompts are in french, do not hesitate to modify them within the file constants.py

Personal notes about the number of token

book : 146.414 tokens Balex une heure : 6000 David une heure 30 : 29000

GPT4 turbo : 32 k tokens max, but doesn't works that well

0.01 centimes pour 1k tokens entrée