Voice debugging at the conversation level seems far more useful than isolated benchmark metrics [D]

I have been thinking a lot about how poorly isolated benchmark metrics capture real conversational system quality once models are deployed into multi-turn environments. You can have strong STT scores, decent latency, high task completion rates, and still end up with conversations that humans perceive as frustrating or unnatural. In practice, many failures are emergent properties of the interaction itself rather than single model errors. Small timing mistakes accumulate. Repeated confirmations create frictio