Very nice idea and great execution. Sometimes it includes clips where the generated subtitles are wrong and the word is actually a different one with similar pronunciation (German heiß -> heißt for a lot of this). Which brings to mind an interesting bias where it leaves out any examples that the AI transcription didn't recognise as the word, thus presenting only the "canonical" pronunciation according to whatever process trained the AI and potentially propagating AI artifacts into the speech of actual humans.