I’m looking at OpenAI’s audio tools for transcribing recorded meetings, and speaker labels are where I’m unsure what to expect. In clean one-person-at-a-time audio, diarization seems straightforward; overlapping speech is a different story.

Has anyone tested how well the API keeps speakers separated when people interrupt each other or speak briefly in the background? I’m especially interested in whether the labels stay consistent through a longer recording, rather than just looking right in a short sample.