Since you are using LLMs to create the transcriptions, I wonder whether you've measured the difference in precision between the chosen model, Gemini 2.5 Flash-Lite, and newer/larger models such as Gemini 3.5 Flash, Gemini 3.1 Flash-Lite or GPT5.5.
I've read the README in the feat-api branch and, from what I understand, you've already assessed that false negatives are not a model failure, but I'm not sure I understand why (haven't spent that much time looking at it though, just curious to hear from you).
This is a really cool project, by the way! In my opinion this is a place where LLMs shine: produce the work of hundreds of hours of manual human labor much quicker and cheaper, for something that no one else would ever bother to do the work!
I've read the README in the feat-api branch and, from what I understand, you've already assessed that false negatives are not a model failure, but I'm not sure I understand why (haven't spent that much time looking at it though, just curious to hear from you).
This is a really cool project, by the way! In my opinion this is a place where LLMs shine: produce the work of hundreds of hours of manual human labor much quicker and cheaper, for something that no one else would ever bother to do the work!