Thanks for sharing. This is exactly the type of utility that vibecoding is for. It takes 5 secons to ask GPT to write a scripr to do this tailored to your specific use case. It's way faster than trying to get someone elses repo up and running.
Youtube already offers AI transcriptions on their site. As another commenter points out, you grab them with yt-dlp.
And unlike how your tool will be supported in the future, thousands of users make sure yt-dlp keeps working as google keep changing the site (currently 1459 contributors).
Many channels I follow, such as Vlad Vexler, have taken measures so you can't download the transcript with yt-dlp. Furthermore, they don't provide a transcipt option on their videos. I assume this is to prevent people from just reading AI summaries, which is annoying in Vexler's case because he talks slowly and meanders around. If I really want to hear his point but don't want to listen to that then I download the video with yt-dlp and use Whisper to transcribe it.
I'd be really curious to see some sort of benchmark / evaluation of these context resources against the same coding tasks. Right now, the instructions all sound so prescriptive and authoritative, yet is really hard to evaluation their effectiveness.
I tried it on a M1 Pro MBP using Docker. It's quite slow (no MPS) and there are no timestamps in the resulting transcript. But the basics are there. Truncated output:
Fetching video metadata...
Downloading from YouTube...
Generating transcript using medium model...
=== System Information ===
CPU Cores: 10
CPU Threads: 10
Memory: 15.8GB
PyTorch version: 2.7.1+cpu
PyTorch CUDA available: False
MPS available: False
MPS built: False
Falling back to CPU only
Model stored in: /home/app/.cache/whisper
Loading medium model into CPU...
100%|| 1.42G/1.42G [02:05<00:00, 12.2MiB/s]
Model loaded, transcribing...
Model size: 1457.2MB
Transcription completed in 468.70 seconds
=== Video Metadata ===
Title: 厨师长教你：“酱油炒饭”的家常做法，里面满满的小技巧，包你学会炒饭的最香做法，粒粒分明！
Channel: Chef Wang 美食作家王刚
Upload Date: 20190918
Duration: 5:41
URL: https://www.youtube.com/watch?v=1Q-5eIBfBDQ
=== Transcript ===
哈喽大家好我是王刚本期视频我跟大家分享...
23 comments
[ 4.2 ms ] story [ 39.6 ms ] threadI must say, speaker diarization is surprisingly tricky to do. The most common approach seems to be to use pyannote, but the quality is not amazing...
yt-dlp --write-auto-subs --skip-download "https://www.youtube.com/watch?v=7xTGNNLPyMI"
And unlike how your tool will be supported in the future, thousands of users make sure yt-dlp keeps working as google keep changing the site (currently 1459 contributors).
- This python one is more amenable to modding into your own custom tool: https://hw.leftium.com/#/item/44353447
- Another bash script: https://hw.leftium.com/#/item/41473379
---
They all seem to be built on top of:
- yt-dlp to download video
- whisper for transcription
- ffmpeg for audio/video extraction/processing
https://github.com/Dicklesworthstone/bulk_transcribe_youtube...
I ended up turning a beefed up version of it which makes polished written documents from the raw transcript, you can try it at
https://youtubetranscriptoptimizer.com/
https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2
For Apple Silicon (MLX) https://huggingface.co/senstella/parakeet-tdt-0.6b-v2-mlx
I'd be really curious to see some sort of benchmark / evaluation of these context resources against the same coding tasks. Right now, the instructions all sound so prescriptive and authoritative, yet is really hard to evaluation their effectiveness.
Uses yt-dlp, whisper, and a LLM (Gemini hardcoded because it handles long contexts well, but easy to switch) for summarizer.
I dislike podcast as a format (S/N level way too low for my taste), so use this whenever I want to get a tldr of some episode.
I should check out the SOTA models and improve the summarization prompt, but aren't in a hurry as this works pretty well for my needs already.