Skip to content

Evan-Wildenhain/CTRL-F-VIDEO

Repository files navigation

CTRL-F for videos

An audio to text word search

How to run

  • import the folder into chrome extensions using the load unpacked function
  • "python app.py" while in the pythons folder this will start a flask API that allows the model to run locally on your gpu.
  • Go to a youtube url, click the extension type in a word or phrase or subword and click mark video.
  • Mark Video - shows direct matches in the color green
  • Show similar shows similar in levenshtein/phoneme distance in the color purple (not that good) and phonetic identical matches in the color orange (occasionally good)
  • Show all - show both green and purple
  • Remove stamps - sometimes it carries over to other youtube pages and just sits on the timebar (only visible when timebar is up), but this is just for those when you want to wipe the colored stamps.

image image image

About CTRL-F

  • The purpose is to create an easy way to search videos for specific words/phrases
  • Before whisper a lot of massaging and similarity searching was required to bring accuracy up, and while implemented they are rarely utilized due to the effectiveness of whisper

Features

  • Audio to text conversion using OpenAI whisper, credit to those at faster-whisper for drastically improving speed over base whisper models.
  • json storage of results for subsequent searches on a url
  • phonetic similarity option (pretty much useless due to the effectiveness of whisper)
  • timebar highlight of location in video on youtube

Bugs/Limitations

  • Only works for youtube as of now
  • First search will not highlight the timebar while the model is running if you leave the page (simply search it again and the cached results will be instant)

Efficiency

  • RTX 3090 can do a 5 minute video in approximately 16 seconds

Feel free to take the project make it your own, this was just a small side project for me.

About

Ctrl-f for videos

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published