Translating a Zoom meeting, and what it costs you in delay

Three options, three failure modes

Live captions. Zoom will caption and translate. It is the cheapest option and it fails quietly: a mistranslated sentence looks exactly like a correct one, and nobody in the call can tell.

A human interpreter. Accurate, and the meeting takes roughly twice as long because everything is said twice. Worth it for a contract negotiation, hard to justify for a weekly standup.

A speech-to-speech tool. The machine speaks the translation out loud. Fast enough to hold a conversation, and wrong in ways that sound confident.

Where the delay comes from

Three things stack: waiting for the speaker to finish a phrase, the translation itself, and generating the audio. The first is the one people underestimate. A system that waits for a sentence boundary is more accurate and feels slower; one that starts early feels responsive and revises itself mid-sentence.

Streaming the speech synthesis instead of waiting for the whole utterance takes a noticeable bite out of the total — that is most of the difference between a tool that feels usable and one that does not.

Why Vietnamese is harder

Tone carries meaning, so a recogniser that drops tone marks produces a word that is a real word and the wrong one. Vietnamese also has less training data than English or Chinese, and most models were tuned on read speech rather than three people talking over each other on a bad connection.

What to do with it afterwards

The transcript is worth more than the live translation. Send it round after the call with the decisions marked, and anyone who missed a sentence can read it instead of asking. That is what the meeting assistant is for; live translation is for the conversation itself.