A live interpreter for Zoom, Google Meet and calls: speak your own language

The short answer

A live interpreter for calls is not a dictionary problem, it is a latency problem. Translating a sentence is something anything can do today. Doing it so the conversation stays a conversation, rather than correspondence through a middleman, is something far fewer tools manage.

There are exactly two working approaches, and you choose between them by whose call it is.

• A call in somebody else's service — Zoom, Google Meet, Teams, Discord, anything. Then the interpreter lives on your computer, listens to the microphone and the system audio, and speaks into your headphones. Nobody has to be invited into the call.
• A meeting you host yourself — then it is simpler to turn on subtitles with translation right in the room: everyone reads in their own language and nobody has to install anything.

What follows: why the familiar «translate in the next tab» approach destroys a conversation, where the delay comes from, how much of it is acceptable, and what a scenario looks like when you speak your own language and are understood.

Why ordinary translation fails in a live conversation

A translator in the next tab looks free right up to the first international meeting. Count the loop: hear the sentence → switch windows → type or paste → read the translation → compose an answer → switch back → speak. Even for a trained person that is ten to fifteen seconds per turn. In a dialogue where a normal gap between turns is under a second, that is an eternity.

Worse is something else. During all that time you are not in the conversation: you are operating the translator. The other side sees a person who looks away, goes quiet and types. They draw the only available conclusion — that the question was hard or unpleasant.

The last line is the important one. A person forced to speak a foreign language keeps sentences short, drops the qualifications and almost never argues. In a negotiation that costs exactly as much as the concession you could not explain.

How live interpreting differs from translating text

The difference is not quality. It is when the decision has to be made.

Text translation works with a finished whole. It has the full sentence, the full paragraph, sometimes the full document. It sees the end before it translates the beginning, so it can reorder words, pick a term and re-read itself.

Live interpreting works with a stream. The end of the sentence does not exist yet, and you have to speak now. That is exactly why human interpreters work with a few seconds of lag and in pairs of half an hour: holding somebody else's thought while uttering the previous one is hard labour, not vocabulary.

Why latency decides everything

Dialogue runs on turn-taking. A pause of up to a second reads as «they are thinking»; two or three seconds read as «bad connection»; more, and the other person starts speaking and you talk over each other. That is why an interpreter five seconds behind does not save the conversation, it breaks it: technically it works, practically people switch to English out of politeness.

The delay has two parts, and they are worth keeping apart.

1. Waiting for the end of the sentence. The system has to decide you have finished. Usually that is a silence threshold: quiet for N milliseconds means the turn is over. Below two hundred milliseconds is dangerous — a short pause inside a sentence («I think… that here») starts being taken for its end, and the translation falls apart into pieces.
2. The model's time to first sound. This is the price of processing itself, and it is several times larger than the first part. It shrinks not by tuning the threshold but by changing the shape of the work: not «receive the whole sentence → translate → return», but start emitting the translation as it appears.

The difference between the two shapes is measurable. We measured it on the same real speech on 3 August 2026: a request carrying the whole phrase returned a subtitle after 4032 ms; streaming the same speech gave the first text after 199 ms. Twenty times — and that is the price not of the model or the plan, but of whether we wait for the end of the sentence.

  WAITING FOR THE END OF THE SENTENCE
  speech     ████████████████
  pause                      ██
  processing                   ████████
  translation                          ███████████████
                             └ the listener is already waiting here

  TRANSLATING AS IT GOES
  speech     ████████████████
  translation   ██████████████████
                └ lag on the order of one phrase

Hence the honest formulation — «practically in real time» rather than «with no delay». Zero latency does not exist even for a human interpreter: they too hear first and speak second. The question is whether the lag fits inside the pause the other person already leaves.

What a good scenario looks like in Zoom, Meet or a call

The requirements follow from the conversation rather than from a feature list. There are five.

Technically it looks like this: your microphone and the sound from your speakers are handled separately, the translation of the other person plays into your headphones, and the translation of your speech goes into a virtual audio cable that Zoom or Meet has selected as its microphone. The virtual cable is an ordinary free utility (BlackHole on macOS, VB-Cable on Windows), installed once.

  THEM → YOU                            YOU → THEM
  audio from Zoom                       your microphone
      ↓                                     ↓
  recognition + translation             recognition + translation
      ↓                                     ↓
  your headphones                       virtual audio cable
  (original ducked)                         ↓
                                        selected as the mic in Zoom
                                            ↓
                                        heard by the other side

  On screen meanwhile: two lines — what was said, and how it was translated.

A live scenario: the client you would rather not lose

Maxim is an engineer and co-owner of a small studio. The client is a manufacturing company in Valencia; three of them are on the call and one speaks confident English. It used to look like this: Maxim prepared phrases in advance, spoke slowly and more briefly than he thought, and answered «why two weeks at this stage and not one» with «because integration is complex» — because the real answer would have taken him a minute in English.

He did not lose on price. He lost because he could not explain.

With live interpreting it looks different. Maxim turns translation on before the call and picks the languages — his Russian and the client's Spanish. At the start of the meeting he says one sentence that is always worth saying: «I will speak Russian, you will hear a translation; if something sounds odd, say so and we will fix it.» Then he speaks normally: with subordinate clauses, with qualifications, with a joke about deadlines. The client hears Spanish roughly a pause behind; two lines run on Maxim's screen — original and translation — and when the conversation reaches numbers and component names he checks them with his eyes.

The outcome: the same two weeks, but now the client understands what they consist of. And after the meeting Maxim has a transcript to write the summary email from, rather than his memory.

Six situations where translation changes the outcome

Why living in window switching is a bad idea

Switching has a measurable price: every return to the conversation costs several seconds of restoring context. In an hour-long meeting with a couple of dozen turns that is no longer a detail — it is half of your participation.

So a proper interpreter is built as a layer over the conversation rather than as a separate application you have to visit: the translation plays in your headphones, the text runs as a line on screen, control is a hotkey. You are not «using a translator» — you are having a conversation.

It also follows that the translation panel should stay out of screen sharing. Not for secrecy but for manners: when you present a slide you present a slide, not your kitchen with two lines of subtitles in it.

What AI actually changed

Five years ago simultaneous interpreting was a profession and a budget line: a human interpreter, materials in advance, agreed terminology, an hourly rate. For a forty-minute call with a prospective client that simply did not pay off, so people spoke bad English and considered it normal.

Three things changed.

• Speech began to be translated into speech directly. The old chain was «recognise → translate text → synthesise voice», and every link added its own delay and its own error. Modern voice models work with audio as a stream.
• Processing became streaming. The answer starts sounding before the question has ended — that same difference between four seconds and two hundred milliseconds.
• The price fell to a subscription. Translation stopped being an event you prepare for and became a toggle you flip before a call.

What AI did not change: responsibility for meaning. Machine translation does not sign contracts, does not take on legal wording and does not replace a written translation where the document will be read in court. And it still loves translating idioms literally — so on an important call speak more plainly than usual. Which, incidentally, helps even without translation.

How this works in Whisperer

We built translation two ways — one per situation from the opening.

Live voice translation on top of any call. Works in the app on macOS and Windows, over Zoom, Google Meet, Teams, Discord and anything else that plays sound on your computer: no bot joins the call, no integration with the service is required. What is inside:

• two directions at once: you hear their translation, they hear yours;
• two modes to choose from — turn-based (translates after a pause, but lets you pick a male or female voice) and continuous (starts translating without waiting for the end of the sentence, with no voice choice);
• 97 languages in the first mode and 79 in the second — the lists are explicit, and on an unsupported language the system refuses honestly instead of quietly translating into the wrong one;
• subtitles on screen: original and translation as lines, can be turned off;
• translation volume and ducking of the original on sliders, so you are not hearing two speeches at once;
• the return channel (them hearing you translated) needs a virtual audio cable; the app installs it for you at the press of a button, verifying the installer's signature. Without the cable translation works one way and says so plainly;
• if the model provider will not admit your location directly, the client routes through our server on its own — no settings, no questions to you.

Voice translation is part of the Max plan, and minutes come from the shared pool — with honest accounting for the fact that there are two directions and roughly double the spend. Hiding that in small print would be pointless: you would see it on the counter by the second meeting.

Subtitles with translation in a meeting you host. If the call runs in a Whisperer room, participants install nothing: a phrase is recognised once at the speaker and translated into each viewer's language — out of the platform's 16 languages. A guest who joined by invitation link reads subtitles in their own language without an account at all. It runs as a stream rather than «one request per whole phrase» — those same 199 ms to first text against 4032 ms that opened the latency section.

The meeting translator quota: on Start and Pro it is 60 minutes of speech per rolling seven days (a minute counts while people talk, not while the translator is switched on); on Max the spend goes into the shared premium budget.

And what we do not promise. We do not promise zero latency — it does not exist. We do not promise that machine translation replaces a certified translator where one is required. And we do not promise that translation saves a meeting nobody prepared for: it removes the language barrier, not the absence of an agenda.

Translation is one capability of a wider class of tools; what they do besides languages is covered in the pillar piece on AI meeting assistants.

Frequently asked questions

What to do next

Language should not be the reason somebody stays quiet in a meeting. If you understand the subject better than anyone in the room but speak about it worse than anyone, the loser is not only your career — it is the decision made without your argument.

And the main criterion for choosing any tool of this kind: a good interpreter must not become a third participant in the conversation. It goes unnoticed. It does not demand switching, does not make people wait and does not turn a dialogue into a contest with a browser window.

Test that on an ordinary call rather than in a negotiation: create an account, turn translation on for your next meeting with a foreign colleague and watch two things — whether it keeps inside the pause, and whether the conversation still feels like a conversation. If yes, you will stop preparing phrases in advance. If not, you will at least learn your own threshold of acceptable delay — it differs for everyone.