Very cool! How does Lyra handle overlapping speakers or messy real-world conversations - does it still manage to extract clear action items accurately?