How to Benchmark AI Voice-Agent Latency Reliably

Latency measurements for voice systems depend on how the timer is defined, since time to first model event, first audio chunk, audible playback, and full response completion each answer different questions. The guidance recommends measuring the interval from the end of a user turn to the first audible response, while separately tracking first audio chunk and full completion times. Benchmarks should document the device, audio input, network conditions, region, configuration, and whether connections were pre-opened, and should avoid mixing cold and warm turns without labeling. Results should be repeated enough to show variability and reported with sample counts and distributions such as median and p95 rather than the fastest run. Latency should also be checked alongside interruption handling, intelligibility, error rates, and dropped turns, since a fast but incomplete response may not be useful.