The generated timestamps are surprisingly amazing, but I found it can trail off at longer passages.
It’d be nice to trigger a “regenerate timestamps from here” inside a given lesson so that I can help the algorithm know where to start in a given piece of audio, and help it generate better matches to the text provided.
Generally, I’d expect a button to appear on an individual line within the lesson to called “Regenerate timestamps from here” or something akin to that, and surface a warning that it’s recommended that the current line in the lesson be accurately lined up with the text, so that the the algorithm knows where to start within the audio on the next piece of text in the lesson. I’d also surface that the timestamps ahead of that sentence would be deleted.
There could be more sophisticated forms of it, where generally you might have anchor points in the audio that align with pieces of text (i.e. you know the starting points of a chapter in a given piece of audio, or structural point in the content & can ask the algorithm to try to align some known X length of audio, before it performs the next task of trying to align some X length of audio of the next. That’s a bit more intense than the prior recommendation, but would also be helpful!
Importantly, a notification should appear that tells us the process is underway (or in a queue if that’s really what’s going on in the background) or at least some form of a “in process of regenerating timestamps” so that you know when you can come back to the lesson.