Jezper Söderlund@marcoarment with transcription, Don you have everything you need to generate fairly accurate chapters for basically every podcast you’ve transcribed?
Fredrik Björeman@Jezper @marcoarment Very curious to hear the thoughts of someone who's thought more about this. My own experience of basic work with what Apple provides is that they don't really have anything for the chapter detection itself. It can transcribe and generate titles based on the text in a time range and other cool stuff, but the context window is too small to ask it to do anything with a full transcript.
@bjoreman @marcoarment that puts some pressure on my ”everything” there I guess, but with a local model and 48 Mac Mini’s soon through the big batch of transcription, it could start generating summaries with time stamps that could act as chapter files. No 600x perhaps, but like with transcripts, you can start with most popular and work your way back. DAI will perhaps mess with timing. Well, still an interesting idea I think 🙂
@Jezper @marcoarment Yeah, I'm just thinking that the parts I've seen and how I can think of putting them together probably wouldn't reach my definition of "fairly accurate".
Then again, I did not consider the option of using anything other than Apple's builtin stuff, so any third-party thing possible adds a very interesting twist
Dang, I should look! Perhaps someone's even done a small local model trained specifically to detect topic changes in speech. Whole different ballpark in that case!