Conversation runs on milliseconds. Passing a phone takes seconds.
Better Translate Team ·
A conversation is a timing system, and it is a tight one. The largest cross-cultural
study of turn-taking found the same underlying pattern everywhere it looked: talk avoids overlap and
silence between turns gets minimised, in every language in the sample.
That study, published in PNAS in 2009, covered ten languages across five continents, from Danish
to ǂĀkhoe Haiǁom. Cultures differ in the average gap, but only within a range of about 250
milliseconds of the cross-language mean. The variation is quantitative. The system is universal.
What that means for a translated conversation
Now add a phone. Not the translation, the phone: unlocking it, turning it around,
holding it out, waiting while somebody else reads, taking it back. That is not milliseconds. On a
good run it is three or four seconds per turn, every turn.
The industry that builds voice assistants has converged on similar numbers from the other
direction. A widely used working target is a response inside roughly 300 milliseconds, on the
grounds that this is about the natural pause in human conversation, and that beyond it people start
to read the system as broken rather than slow.
Interpreting tolerances split the same way. Conference interpreting has an established tolerance
of roughly three to five seconds, because the audience is passive and used to it. Bilateral
conversation, two people going back and forth, has far less room: practitioners describe a two
second delay as disruptive and four seconds as making the exchange feel broken.
Which is exactly the budget a phone handover spends before any translation happens.
So we stopped spending it
The screen splits down the middle and the far half is rotated 180 degrees. Your half faces you,
theirs faces them, and both are live at the same time. The phone stays flat on the table.
That does not make the translation faster. It removes an entire category of delay that had
nothing to do with translation in the first place. The handover was never the hard engineering
problem, and it was costing more seconds than the model was.
The honest limit
One microphone means one side talks at a time. Turn-taking is preserved, because turn-taking is
how conversation works anyway. What is removed is the passing.