A numbered series on the speech recognition, speaker diarization research and engineering behind Thaki Minutes, a Korean AI meeting-notes system in development.
We trained and released a Korean LoRA for a 30B speech recognition model. What got it through the release gates was not only the training recipe but a decodi...
Switching the diarization method cut wrong merges sharply in larger meetings and made one set of small meetings worse. Here is how we read those numbers, and...