So I've been working on a side project to make a Youtube channel I watch have its content be more discoverable through text. I've had great results by scraping the Youtube transcription, and running that for a few passes through GPT 3.5 with some prompts to essentially act as an editor. The original transcription was often terrible in some spots. Just whole phrases or multiple words mistranscribed throughout. For almost all of them, GPT 3.5 was able to clean them up and restore the original meaning through understanding the context of the monolog and fixing obviously incorrect words or phrases.
I've watched through a sample of about 20 of the 3,000 videos I'm working through, and the corrected transcription really did an amazing job at restoring the original meaning from the spoken words that was hard to understand from the original machine transcription.
That is exactly where LLMs are useful. (People thinking of them as "AI", meaning AGI is just so wrong. Writing legal briefs??) Using them to ex post facto adjust transcripts in order to make them available and searchable is great.
I've watched through a sample of about 20 of the 3,000 videos I'm working through, and the corrected transcription really did an amazing job at restoring the original meaning from the spoken words that was hard to understand from the original machine transcription.