Wrong: "YouTube doesn't have an API for transcripts."
It doesn't have a PUBLIC one. But transcripts exist for nearly every video including auto-generated ones.
~94% of videos have extractable transcripts. The data is there. The access layer was missing.
Wrong: "Auto-generated captions are garbage."
That was true in 2019. Post-2022, YouTube's speech recognition is remarkably accurate.
We've processed millions of auto-generated transcripts. Error rates are low enough for RAG, search, and summarization at scale.