Scribe - Beautiful YouTube Transcripts with Gemini 1.5 Flash 8B LLM

This repo shows how to fetch the raw YouTube transcripts and use the Gemini Flash 8B API to format them. Made by @ldenoue

How to use

Try the demo at https://ldenoue.github.io/readabletranscripts and type any search term

Try locally python3 -m http.server and open http://localhost:8000

Set your Gemini API KEY (create a Free Gemini API Key) Note: The key is used

readabletranscripts/code.js

Line 159 in f5ea580

const genAI = new GoogleGenerativeAI(API_KEY);

to fetch Gemini's answers from your browser. It is never uploaded anywhere.

Click on the video, e.g. https://ldenoue.github.io/readabletranscripts/?id=8yzmCt0QwOQ

You will see a summary of the video and the transcript below.

How it is done

Getting vocabulary words from the title and video description

We first ask Gemini to extract important words by giving it the video title and its description. See

readabletranscripts/code.js

Line 837 in 8466ec2

async function createVocabulary(videoId, description = '', languageCode = 'en') {

This context is essential to improve the accuracy of the transcripts. Titles and descriptions often contain human-edited text that includes proper names, acronyms, etc.

Breaking the transcript into chunks

We break up the raw YouTube transcript into chunks of 512 words. We feed each chunk to Gemini with a prompt (see

readabletranscripts/code.js

Line 410 in 5c51f52

async function punctuateText(c, vocab = '', lang = 'en', p = null) {

)

Notice that we send the requests in parallel to Gemini.

Merging sentences at the boundary between 2 chunks

Once we have the formatted chunks, we now need to merge them. For each 2 consecutive chunks chunk1 and chunk2, we ask Gemini to merge the last sentence chunk1 and the first sentence of chunk2 (see prompt in

readabletranscripts/code.js

Line 433 in 5c51f52

async function mergeSentences(a, b, vocab, languageCode = 'en') {

)

We merge the chunks and the seams between them to obtain the final transcript.

Linking the raw YouTube timestamps with the final transcript

In order to highlight the words as the video plays, we need to align the words from the raw YouTube transcript and the final, punctuated, transcript.

We rely on diff.js for that.

Now words get highlighted as the video plays, and users can also jump into the video by clicking any word in the transcript.

Help me improve this tool

Thanks for your contribution.