Streaming recognition, incremental generation, interruption handling and speculative response starts
An agent that is accurate and slow feels broken to a caller; one that is fast and occasionally asks for clarification feels normal
We tune against recordings of your actual calls rather than a benchmark








