How far behind the speaker AI interpretation runs: what we measured in a live room

In the main hall of a Toronto forum in October 2026, the Chinese voice started a median 2.9 seconds after the speaker, and within 4.4 seconds for 95% of sentences. How it was measured, when it runs later, and whether captions or voice come first.

What is measured

The time from a speaker starting a sentence to the first audible interpretation of that sentence leaving our server. Silence the engine puts at the start of its voice does not count: the clock stops at the first frame that can be heard. The figure leaves out the network to the phone and the phone’s playback buffer.

In a live room

Main hall of the My Healthcare Forum, Toronto, 29 October 2026, English on stage, Chinese channel: of the 16 sentences that could be measured, a median of 2.9 seconds, and 95% within 4.4 seconds. A small sample, for that one event.

When it runs later

When a speaker goes a long time without a pause, the voice falls further behind: the engine has to hear enough to translate the meaning. We compared on an earlier engine version: short conversational sentences, 95% within 3.7 seconds; long passages read aloud, 95% within 5.6 seconds.

So the one piece of advice for speakers: pause briefly between sentences.

Captions or voice first?

Almost together. On our test recordings, captions led the voice by a median of 0.15 to 0.25 seconds, and never came after it. The big screen and the phones show the same captions.

Why the landing page says two seconds

In the landing page samples, the voice starts about two seconds behind the stage. The default sample is a clear, steady solo talk (all of the landing-page samples except Mandarin are real recordings); a live room has pauses and long sentences, hence the median of 2.9 seconds above.

Read next

Try it free for 5 minutes

Markdown