I wonder if they couldn't have solved that in some technical manner too. For example have the autogen transcripts first by default, but allow the opportunity fo

I wonder if they couldn't have solved that in some technical manner too. For example have the autogen transcripts first by default, but allow the opportunity for community contributions to first curate the autogen transcripts and correct any devations where the autogen is wrong. So we start with preclassified speech and correct that small fraction of the time the autogen is mistaken. I figured they could have even used those delta transcripts to improve training on. Videos with huge error rates could flag the contribution for human review, and even with those ones that faithfully correct have some human review them quickly since it may be a word or two. I usually catch at least one discrepancy every video I watch with captions that's longer than say 10 minutes. More frequently on technical subject matter too.