https://api.github.com/repos/Dannydoesdev/summarise-and-shine/contents/{+path}

cute robot typing in computer realistc.jpg

MIT License

summarise-and-shine

Summarise and Shine

<h1 style="text-align: center">YT Summariser</h1>YouTube Summary with Whisper API, ChatGPT API + Eleven Labs Voice<h2>Why this?</h2>I joined the DSAI GPT-4 hackathon (<a href="https://thechainsaw.com/business/chatgpt-hackathon-sydney-ai-noob-what-we-learned/?itm_source=wp-parsely-recommendations-block" rel="noopener noreferrer nofollow" target="_blank">more about it here</a>) with a couple of founders who were doing Startmate's Launch Club program with me at the time. We teamed up with a couple of new faces to jam on a problem we could try to tackle over a weekend using AIBeing the only dev on the team and also being wildly (read 'naively') ambitious, we attempted to create a kind of 'second brain' for content.We had access to Azure's AI cloud system and a bunch of credits to try and use - so I stayed up most of the weekend trying to create what would basically be 2 seperate but linked solutions. A private ChatGPT trained on content you followed (podcasts/youtube/articles etc) and to feed that system, a summariser bot that would create various snippets of content.This is the second part - the first actually got made too, but without the access to Azure any more can't really be shared.<h2>What was learned?</h2><ul><li>Hackathons are awesome - and wow people can do a lot in one weekend when there's the 'right amount' of constraints</li><li>Building a thing and 'selling' people on that thing are very different, we didn't make it to the finals even though I still think our solution was pretty awesome. I think it was hard to actually pitch it when we all had slightly different ideas about what 'it' was!</li></ul><h2>Purpose</h2><ul><li>To let users consume content in the forms they want, not necessarily the forms they are uploaded in</li><li>To reclaim time spent watching long form content when not necessary, summarisation can help understand whether longer content is worth your time</li><li>Proof of concept for a larger suite of personalised content summarisation tools</li><li>Learning more about user cases for AI &amp; the ease of developing helpful micro tools with ChatGPT</li></ul><h2>What it does</h2><ul><li>Takes a link (currently a Youtube video link only), transcribes the content then loops over the transcription a few times to generate summaries of different lengths - including a long bullet points summary and a short paragraph summary. (<a href="https://drive.google.com/file/d/1r3rSzk7xEBOIBbsLLA_ecEmQ_NengpLm/view?usp=drive_link" rel="noopener noreferrer nofollow" target="_blank">Example here</a>)</li><li>Uses whisper API to turn the short summary into a ~1 min explainer you can listen to - (<a href="https://drive.google.com/file/d/1yvA6yAG3p6hbH5EeUezpVG3r-XFyAVSk/view?usp=drive_link" rel="noopener noreferrer nofollow" target="_blank">Example here</a>)</li><li>Outputs all of this to a folder within 1-2 minutes total</li></ul><h2>Personalisation</h2><ul><li>You are welcome to try it out - it will be easiest for those with some coding knowledge</li><li>The current prompts are just experiments and you are encouraged to try variations to find something that works for you</li><li>systembot.txt will set a system message for the gpt functions - leaving this blank has worked well so far</li></ul><h2>How it works</h2>Note: All summarised files are placed in a './summary/' subfolder<ol><li>Downloads high quality version of youtube video via pytube</li><li>Extracts audio into segment files via pydub so it can be transcribed to text via Whisper API</li><li>Uses Whisper API to transcribe segments</li><li>Joins segments back up - full transcription can be found in 'podscript.txt'</li><li>'podscript.txt' fed to GPT in chunks with the task in 'prompt1.txt' to make a full summary, summarised to 'initialsummary.txt'</li><li>'initialsummary.txt' is broken into chunks and put through GPT with the task 'prompt2.txt' to make bulletpoints and 'prompt23txt' to make a concise summary, summaries saved in 'bulletpoints.txt' and 'shortsummary.txt'</li><li>'bulletpoints.txt' fed into gpt to generate a narration synopsis - 'synopsis.txt'</li><li>'synopsis.txt' is sent to elevenlabs API to be transcribed to an audio spoken file - audiosynopsis.mp3</li></ol><h2 style="text-align: start">Running in colab</h2><ol><li>Open colab file from <a href="https://colab.research.google.com/" rel="nofollow" target="_blank">colab</a> - you likely need to create a new notebook and upload to your drive</li><li>Browse to colab hosted resources from the folder icon on the left pane</li><li>Upload all necessary local files (can copy contents of the 'copy-my-contents' folder)</li><li>Set your keys in the openai and elevenlabs txt files - note this is not safe to share, feel free to secure colab api keys in other ways</li><li>Set intended Youtube URL in "URL.txt"</li><li>Run colab sections</li></ol><h2></h2>Extra bonus content - <a href="https://docs.google.com/presentation/d/1qTdd6n5GxRL9MXhZ92FON9oatw-0d3DnMbZ2L6-I9po/edit?usp=sharing" rel="noopener noreferrer nofollow">our pitch deck for the hackathon</a>