This site is a sandbox project to try out Whisper by OpenAI.
Whisper is a machine learning powered speech to text engine that is fairly easy to use and impressively accurate.
In order to make learning a new thing "stick", IMO, it's always best to wrap it in something personally interesting if you can.
For example: "I'm going to learn how this speech to text AI thing works, so I'm going to feed it all the JB content I like to listen
to and see what pops out!"
This project also gave me a chance to warm back up with django, and see what results full text searching with
django-Haystack and Whoosh would produce.
Whisper outputs in TXT (no time codes), SRT (SubRip subtitle format), and VTT (video text track) formats. The VTT files can be used with audio/video players to give you a live close caption or live transcript that follows along as it plays. This demo site uses AblePlayer.
Note: Content presented was batch processed while I had temporary access to an Nvidia 3090 and leveraged the "large" model with Whisper. Files can be processed with CPU only, but takes hours instead of minutes to process one show. Automating the processing of new episodes is something I'm interested in, but gets a bit outside of the scope of the original "learning whisper" goal and is a version 2.0 type of feature. I also figured it's worth throwing this over the fence to the community just in case it's something that others more familiar with JB content would want to look into.
Transcript sources can be found at https://github.com/ShuffleBox/JB_content/.
This site source can be found at https://github.com/ShuffleBox/JB_demosite.
Janky scripts to process things can be found at https://github.com/ShuffleBox/JB_M04r.