The One-Second Rule: Why Latency Makes or Breaks Voice AI

Humans take turns in conversation with astonishing speed — typically around 200 milliseconds between one person finishing and the next beginning. We are exquisitely sensitive to delay. That's why a voice assistant that takes three seconds to respond doesn't feel slow; it feels broken.

Why a second is the threshold

Under roughly a second, a response feels conversational. Past it, the human brain registers an awkward gap, the speaker wonders if they were heard, and they start to repeat themselves — which collides with the late response and derails the exchange. In a public space, that awkwardness is amplified by an audience.

Where the milliseconds go

• Speech capture and endpointing — detecting when the user has actually finished speaking.
• Transcription — turning audio into text.
• Retrieval — finding the right grounding content (RAG).
• Generation — composing the answer.
• Speech synthesis — turning the answer back into natural audio.

Every stage adds latency, and they compound. Hitting sub-second end-to-end means engineering each stage and overlapping them — starting to synthesise the beginning of an answer while the end is still being generated, for instance.

Users don't measure latency in milliseconds; they measure it in awkwardness. The target isn't "fast" — it's "no awkward pause".

Endpointing: the underrated half

Half of perceived latency is knowing when the user has stopped talking. Cut too early and you interrupt; wait too long and you feel sluggish. Good systems use natural turn-taking cues and allow barge-in so users can interrupt — the same flexibility people expect from each other.

What to test

Measure end-to-end response time under realistic load and network conditions, in your noisiest environment, across your languages. A platform that's snappy in a quiet English demo and laggy in a crowded multilingual lobby has optimised for the wrong test.