There are auto-generated transcripts on youtube that are searchable. Click the 3 dots at the bottom right of the video to get to it. But yes, a proper transcript would be even better, since there’s plenty of errors in the automatic one.
BTW one idea would be to use Lin’s transcripts to train a better voice recognition model for my voice, and then auto-generate the transcripts for the new videos. You could even learn a language model to make it more accurate… It would be quite a big project, but it would be pretty amazing to let anyone create a speech recognizer from their own youtube captions!
That sounds like a fantastic project!! I usually have the youtube’s transcripts open next to the video, but it could use some improvements. I was quite impressed when it spelled “Leslie Smith” correctly, but other terms like “ReLU”, not so much.
I am aware of your awesome set of notes on all the new lessons including Machine Learning. Really helpful. I have been reading some of those DL Part 2 2018 notes before the official launch. I noticed the notes are pretty much in sync with the video. Easy to follow. So, yeah, in some way, almost like “transcripts” Thank you very much.
Yes, there’s a wealth of content in each lecture. Especially in part 2 where we end up implementing recent SOTA papers/techniques in each lecture. Taking notes and reviewing multiple times is the way to go!
Right. A previous participant have also created a simple Python script to grab the auto-generated transcripts on all fast.ai YouTube videos but the transcription quality is so low that it’s even hard to manually correct the mistakes. But with the part 2 2018 edition, the video quality is so much better, so the auto-gen YouTube subtitles are better: https://github.com/mingrui/youtube_caption_search
Thanks for giving me the idea. Yeah, it’s an interesting idea to turn into a project with great potentials.