Seven checks before you trust a live translator in a meeting
Seven checks to run before you rely on live speech translation in a client meeting, from names and numbers to overlapping speakers and privacy.
Key takeaways
- Test on your own people, your own audio and your own vocabulary, not on a clean demo sentence.
- Names, numbers and overlapping speakers are where live translation fails first.
- Ask what the tool does when it is unsure, and where your audio goes.
- Keep a written recap as the safety net, whatever the result.
A live translator is ready for a client meeting only if it passes your own test, on your own audio, with the people and words that matter to you. A polished demo does not count. The seven checks below take about twenty minutes and need two colleagues, a printed script and the tool you want to judge.
Run them in order. Write down what happened, not what you expected. If you need a pass mark, decide it before you start.
Check 1: does it get the names right?
Put five names in the script: two people, a company, a product and a place. Use names from your real client list, including at least one that is not common in the language of the speaker. Names are the first thing a translator gets wrong, because they are not in any dictionary, and the mistake can be invisible to the listener.
Pass: every name arrives intact, or the tool shows a clear sign that it was unsure. Fail: a name is silently replaced by a similar-sounding word.
Check 2: does it get the numbers right?
Read out an amount with a currency, a date, a percentage and a phone number, once as digits and once as words. The risk is not a dramatic error. It is "eight thousand" becoming "eighteen thousand", or a decimal comma read as a thousands separator.
Pass: every figure matches, in both forms. Fail: any figure changes. Treat a number error as a stop sign, because it is the one that costs money.
Check 3: what happens when two people talk at once?
Ask two colleagues to speak over each other for ten seconds, as people do when they get excited. Then watch what the listener receives. Overlapping speech is hard for people and harder for software, and Google's own help page warns that live translation can have more errors than a recording.1
Pass: the tool keeps the speakers apart, or it clearly drops one rather than mixing them. Fail: the output blends both into one confident, wrong sentence.
Check 4: does it cope with real voices?
Do not test only with a clear studio voice. Use the accents, speeds and microphones your participants have: a laptop microphone in a kitchen, a phone on a train platform, a speaker who talks quickly when nervous. A tool that works for the person who built it may not work for your client.
Pass: the meaning survives with your least-ideal speaker. Fail: it only works when the voice is slow, close to the microphone and in a quiet room.
Check 5: how long is the delay, and can people live with it?
Do not trust a number on a website. Measure on your own call: say a sentence, stop, and count until the listener hears the translation. Then hold a short real conversation, with questions and answers, and ask the listener how it felt. A delay that looks acceptable in a table can feel awkward when you are waiting for an answer.
Pass: the people on the call say they could talk naturally. Fail: people begin talking over each other because they cannot tell whether the other side has finished.
Check 6: what does the tool do when it is unsure?
Mumble a sentence, use a word from an industry few people know, or switch language for half a line. A good tool shows doubt: a marker, a gap, a note that it did not catch something. A bad tool always sounds equally sure.
Pass: uncertainty is visible to the listener. Fail: wrong output looks exactly like right output.
Check 7: where does the audio go, and who keeps it?
Ask three questions and get the answers in writing. Where is the audio processed? How long is it stored? Who can read the transcript? For client meetings, you may owe your client an answer to the same three questions. If the supplier cannot answer them clearly, the tool is not ready for confidential work.
Pass: clear written answers, and the participants know they are being transcribed and translated by AI. Fail: vague answers, or participants who do not know.
How should you score the seven checks?
Give each check a pass, a partial or a fail, and decide in advance what blocks a client call. Our suggestion: a fail on names, numbers or privacy is a block. A fail on accents or delay means you can use the tool for internal calls first and learn from them.
Then repeat the test whenever something changes: a new language pair, a new meeting tool, a new participant with a different voice. Results for one setup do not carry over to another.
Whatever the score, keep a human safety net. Send a short written recap with names, numbers and decisions after every call, and ask each participant to confirm it.
Where does VocalSpan stand on these checks?
We built VocalSpan as a meeting room where everyone speaks their own language and hears the others in theirs, and we run these checks on it ourselves. It is early software. So far we have tested Italian and English in a remote call on 23 October 2026, and we have not tested other pairs. You can see it on your own kind of meeting in a 15-minute call, joined from your own device.
Frequently asked questions
How long does the scorecard take?
About twenty minutes with two colleagues, a prepared script and the tool you want to judge.
What counts as a pass?
You decide before you start. A good rule is that any failure on names, numbers or privacy means the tool is not ready for a client call, whatever else it does well.
Sources and notes
-
Google Meet Help, "Speech translation", support.google.com/meet/answer/16221730 (read 2 October 2026; the wording can change, so check the page before you quote it). ↩