Google has four different translation tools, and they do not do the same thing. The Translate app translates what you hear through your headphones. Google Meet translates everyone in a meeting into each listener’s own language.
The Gemini app interprets a conversation in turns. And behind the newest two sits a single engine, Gemini 3.5 Live Translate, that translates speech to speech in one continuous stream. This guide separates the four, explains how the new engine works, and ends with which tool fits which situation.
The headline change is simple: translation used to wait for you to finish speaking, then translate, then speak. The new engine does not wait. It listens and translates at the same time, trailing the speaker by only a few seconds, and it tries to keep the speaker’s own voice — their pace, their emphasis, their urgency — instead of replacing it with a flat robotic voice.
Four Google Translators, Not One
Most confusion about Google translation comes from the names. “Google Translate”, “Gemini Live” and “Meet translation” sound like one product, but they are four surfaces running on different machinery.
First is the Google Translate app on Android and iPhone. It is the consumer tool: point it at text, type, speak, or listen live. Its live translation feature turns spoken language into translated audio in real time.
Second is Google Meet’s speech translation. It lives inside video meetings and translates participants for each other. Since October 2026 it covers 63 languages, with each listener choosing their own language.
Third is the Gemini app’s interpreter. You talk to the Gemini assistant and it translates back and forth with you. It works, but it works in turns: it listens, waits for a pause, then responds. That rhythm suits a chat, not a live lecture.
Fourth is the engine underneath the first two: Gemini 3.5 Live Translate. It is the model that powers live translation in the Translate app and in Meet, and developers can also use it through the Gemini Live API. When this guide says “the new engine”, it means this model.
Keeping the four apart matters because their limits differ. A feature available in Meet may not exist in the Translate app, and the Gemini assistant’s turn-taking has nothing to do with the engine’s streaming. Compare the surface you will actually use, not the brand name above it.
The Old Way: Three Hops and Waiting
To understand what changed, it helps to see what translation used to do. Classic voice translation ran as a chain of three separate steps.
First, a speech recognition model turned audio into text. Then a translation model turned that text into text in another language. Then a speech synthesis model read the translated text aloud in a generic voice. Engineers call this a cascaded pipeline: speech to text, text to text, text to speech.
Each hop cost time and quality. The system had to wait for a sentence to finish before it could start translating it, so conversations moved in blocks with pauses between them. And each hop threw away information: the original speaker’s tone, hesitation and emphasis never survived into the final voice, because the middle step only carried words.
The result was usable but unmistakably mechanical. You could understand what was said, but not how it was said — whether the speaker was calm or urgent, joking or serious. For a train announcement that hardly matters. For a negotiation or a doctor’s appointment, it matters a great deal.
The New Engine: Gemini 3.5 Live Translate
Gemini 3.5 Live Translate, launched in June 2026, collapses the three hops into one. It is a native audio-to-audio model: it takes in spoken audio and produces translated spoken audio directly, without stopping at text in between.
Because there is no waiting for sentence boundaries, it streams. It follows the speaker continuously and delivers translation a few seconds behind them, rather than in blocks after pauses. Overlapping speech and rapid talkers still challenge it, but the rhythm is much closer to a real conversation than the old turn-based pattern.
The second change is voice preservation. The engine tries to keep the original speaker’s intonation, pacing and emotional cadence in the translated output, so a calm explanation still sounds calm and an urgent warning still sounds urgent. Google says generated audio from its models carries an inaudible SynthID watermark, a marker that identifies the audio as AI-generated without changing how it sounds.
Coverage differs by surface. Google puts live translation in the Translate app at more than 70 languages, around 2,000 language pairs. Meet, which runs the same engine inside meetings, covers 63 languages as of October 2026. The numbers differ because the products rolled the engine out at different times and with different testing, not because the engine itself changes.
Google Translate App: Listening and Conversation Modes
In the Translate app, live translation is a single feature with three modes. You open the app, tap live translate, choose your languages, and pick how you want to hear the result.
Listening mode is for one-way listening: a lecture, a tour guide, a speech, a film. You connect headphones, or hold the phone to your ear like a telephone call, and translations play privately into your ear as the speaker talks. On Android, Google added the hold-to-ear option so you do not need headphones at all in some situations.
Conversation mode is for two-way talk. Translations play out loud on the phone’s speaker, or into headphones, and the app takes turns as each person speaks. A face-to-face mode splits the screen so each speaker can read their own language on their half.
Text-only mode skips audio entirely and shows translations as text. Custom settings let you choose between text, speaker or headphones for each side of the conversation.
Two practical details make the app pleasant to use. The microphone detects automatically when one language stops and another starts, so you do not need to tap between speakers. And on Android, a session keeps running in the background if you switch apps, minimise the app or lock the screen, so a long lecture does not end because you checked a message.
Availability varies. Live translation needs the latest app version, and features differ by region and language. Some options grey out when no headphones are connected, since listening mode without them only works through the phone’s earpiece on supported devices.
For translating printed text rather than speech — a sign, a menu, a document — the camera route belongs to Lens rather than live translation. Our guide to how Lens, Images and Ask Photos work explains that side of Google’s visual translation in detail.
Google Meet: 63 Languages Since October 2026
Meet’s speech translation changed fundamentally on October 22, 2026. Before that update, Meet supported a small set of language pairs — English against five European languages — and a whole meeting had to share one language pair chosen in advance.
The October update, powered by Gemini 3.5 Live Translate, expanded support to 63 languages across web, Android and iOS, with listening support on Meet hardware devices. Any combination of the 63 languages can now be used in the same meeting.
The bigger change is control. Translation is no longer a meeting-wide setting. Each participant turns it on for themselves from the toolbar, chooses the language they want to hear, and can mark languages they already understand so those are left untranslated. If a speaker switches languages mid-meeting, Meet detects the switch and keeps translating.
Captions follow the translated audio, so listeners get a matching reading and listening experience. Speakers choose whether their own speech is translated at all: nobody’s voice is translated without their say-so.
Access depends on plan. Eligible plans include Google AI Pro and Ultra, several Workspace business and enterprise tiers, and education plans. In a group meeting the host’s plan covers guests; in a one-to-one call, either side’s eligible plan covers both. Base plans and Google AI Pro allow up to 90 minutes per meeting, while higher tiers allow up to eight hours.
There are exclusions worth knowing before an important call. Translation does not work in live streams, recordings, client-side encrypted meetings, dial-in connections or calls through third-party interoperability. Meet hardware in a conference room can play remote participants’ translated speech, but it cannot translate speech coming from the room itself.
Gemini App Versus the Dedicated Engine
The Gemini app’s interpreter and the dedicated Live Translate engine look similar on the surface, but they behave differently in practice. The difference is turns versus stream.
The Gemini assistant holds a conversation. It listens until you pause, then it thinks, then it answers. That loop is natural for asking questions and getting explanations, and translation inside it works the same way: you speak, it translates, you hear the result, then you speak again.
The dedicated engine never takes a turn. It translates continuously while people talk, which is what a lecture, a meeting or a live broadcast needs. If you tried to follow a speech through the assistant’s turn-taking, you would fall behind within a minute, because every pause for translation is a pause in the speech you are missing.
The practical rule is about who is talking. If you are talking with the assistant, use the assistant. If you are listening to other people talk — a speaker, a meeting, a tour — use the Translate app or Meet, which run the streaming engine. Choosing the wrong one is the most common reason live translation feels broken when it is actually just mismatched.
Which Should You Use?
For travel and everyday listening, use the Translate app in listening mode. Headphones in, or phone to ear, and a guide, a lecture or a stranger’s directions become understandable without interrupting anyone. It is the lowest-friction option and it works with any headphones.
For a two-person conversation, use the Translate app in conversation mode, or face-to-face mode if you are standing together. Each person speaks naturally and hears the other in their own language. Keep sentences reasonably short; even streaming translation copes better with pauses than with a five-minute monologue.
For a work meeting, use Meet’s speech translation. It is the only option that gives every participant their own language choice in the same call, with captions to match. Check the plan and the 90-minute limit before a long session, and test it on a low-stakes call first.
For asking questions about something in another language — what does this sign mean, how do I say this — use the Gemini app. Its turn-taking is a strength there, because you want an answer and an explanation, not a running translation of a speech.
For building translation into your own product, use the Gemini Live API, which exposes the same streaming engine to developers. That is the route for support tools, meeting products and classroom apps that need live translation without building their own speech models.
Where It Still Falls Short
Live translation is genuinely better than the old three-hop chain, but it is not magic, and its limits should shape how you rely on it.
First, there is lag. Translated speech arrives a few seconds behind the original, and captions can lag further. In a fast back-and-forth, that delay changes the rhythm: you need to leave a beat before replying, or you will talk over the translation still playing.
Second, detection can err. Google’s own guidance warns of occasional language-detection errors and caption delays. Similar-sounding languages, heavy accents and noisy rooms make mistakes more likely. If a translation suddenly stops making sense, the likeliest cause is a detection slip, not a vocabulary gap.
Third, availability is uneven. Features roll out by region, language and plan, and mobile users may need to update the app before the expanded Meet languages appear. A feature your colleague has may not have reached your account yet.
Fourth, some things are simply out of scope. Recordings, live streams and encrypted meetings are excluded from Meet translation, and hardware rooms can only listen, not broadcast translation. For contracts, medical decisions or legal statements, a written record agreed by both sides still beats any live translation, however good.
The honest summary is that live translation has moved from “understandable with effort” to “usable in real conversations”, with a few seconds of delay and occasional errors as the remaining price. For travel, meetings and lectures that is a bargain. For anything where a single mistranslated word changes the outcome, keep a human in the loop.
How Live Translation Handles the Hard Cases
Real conversations are messier than demos. People switch languages mid-sentence, talk over each other, mumble numbers, and hold meetings in echoing rooms. How the streaming engine copes with those cases decides whether it is genuinely useful or merely impressive.
Language switching is the case the October 2026 Meet update was built for. Previously a meeting picked one pair and stuck to it, so a colleague switching from Spanish to English mid-call broke the translation. Now Meet detects the language being spoken and translates into each listener’s preference, even when speakers switch. The Translate app’s microphone does the same within a session, detecting when one language stops and another starts.
Detection is not perfect, and Google says so itself, warning of occasional detection errors. Two languages that sound alike, a speaker with a strong regional accent, or a sentence that starts in one language and ends in another can all confuse the detector.
When the output suddenly turns to nonsense, switching the expected language manually, or asking the speaker to repeat the sentence, usually fixes it faster than waiting for the system to recover on its own.
Numbers, names and addresses deserve special care. A mistranslated verb is usually obvious from context; a mistranslated dose, price or street number may not be. In any conversation where a number changes the outcome, confirm it back: repeat the figure in the original language if you can, or check it in writing before acting on it.
Noise is the other great enemy. A quiet room gives the engine clean audio to work with. A busy street, a car, or a meeting room with several side conversations gives it a mixture, and the translation degrades with the audio. Headphones help twice over: they keep the translated audio private and they keep your microphone closer to the voice that matters.
Long monologues test patience more than accuracy. The engine keeps translating, but the few seconds of lag accumulate into a feeling of being behind, especially if the speaker never pauses. Short turns with natural pauses feel dramatically better, which is why the best live-translation experiences are conversations, not broadcasts. If you are presenting, pause at the end of each point and let the translation catch up before moving on.
Children’s voices, whispering and shouting all sit outside the engine’s comfort zone. None of this is a reason to avoid live translation. It is a reason to use it where its strengths lie — ordinary adult conversation in reasonable conditions — and to keep a written record for anything where precision matters more than speed.
Setting Up Live Translation in Five Minutes
Getting started takes less time than reading this guide. The steps differ slightly by surface, but the shape is the same: update, connect, choose, test.
First, update the app. Live translation features arrive through app updates, and Meet’s expanded languages may not appear until the Meet app is current. Check for updates before an important call, not during it.
Second, connect your audio. For the Translate app, pair any ordinary headphones, or plan to hold the phone to your ear for listening mode. For a meeting, a headset beats laptop speakers: it keeps your translated audio out of the room microphone and stops the system from translating its own output.
Third, choose your languages and mode. In the Translate app, pick the two languages and the mode that matches the situation: listening for one-way, conversation for two-way, text-only if you would rather read. In Meet, open the captions and translation control, choose translate speech, pick your listening language, and mark any languages you already understand so they are left alone.
Fourth, test on something unimportant. A five-minute call with a colleague, or a foreign-language video with the sound on, tells you more about lag, volume and detection on your device than any review. Test the exact headphones and network you will use for the real thing.
Fifth, check the limits that apply to you. Meet’s translation runs up to 90 minutes on base plans and Google AI Pro, and up to eight hours on higher tiers. If your meeting runs longer, plan a break or a second session. And remember the exclusions: no translation in recordings, live streams or encrypted meetings, so if the session must be recorded, arrange a separate written summary.
One last habit pays off quickly: turn captions on alongside translated audio in Meet. Reading while listening catches the errors that hearing alone misses, and the captions give you a written trace of what was actually said when a number or a name matters later.
Background
Google’s translation story runs from typed text in the 2000s, through camera translation and conversation mode in the Translate app, to the Gemini-powered streaming engine that arrived in June 2026 and reached Meet’s 63 languages in October 2026. The direction is consistent: fewer steps between hearing and understanding, and less waiting in between. You can browse the subjects we cover on our topics index.
Sources: Google Workspace Updates: “Meet speech translation”; Google Translate Help: “live translate help”.
0 Comments