Measure response time end to end—not with a “fastest” headline.

The metric: caller stop to audible reply

Use end-to-end turn latency: the elapsed time from the end of the caller’s utterance to the first audible sample of the receptionist’s next response. Do not substitute speech-to-text completion, model first-token time, text-to-speech start, or server-side timestamps that exclude the phone network.

Minimum reproducible protocol

1. Pre-register the vendor list, scripts, locale, device/network conditions, test dates, sample size, and exclusion rules.
2. Use multiple awareness, intake, clarification, interruption, and escalation turns—not one greeting.
3. Run enough repetitions to characterize normal variance; use at least 30 valid turns per scenario and disclose all attempted turns.
4. Record calls only when consent and applicable law permit, and retain synchronized audio or timestamp evidence for review.
5. Separate answer pickup time from conversational turn latency. They measure different experiences.
6. Have an independent reviewer reproduce the calculations and publish corrections when a vendor identifies a factual error.

What a future comparison must disclose

• Who ran and funded the test, including VantaWeb’s ownership interest.
• Whether each system was a public demo, trial, paid account, or custom configuration.
• Exact dates, versions, phone carriers, regions, languages, and time-of-day windows.
• Raw privacy-safe measurements and calculation code.
• Quality outcomes alongside speed; a fast wrong answer is not a win.
• Limitations, missing vendors, confidence intervals, and a vendor correction process.

Only after this evidence exists should VantaWeb describe an observed result, and even then the claim must be scoped to the tested configuration and dates. “Fastest” requires a representative market comparison, not merely a favorable internal number.