You are mid-sentence with someone who does not share your language, and instead of both of you stopping to wait for a phone to catch up, the translation arrives in their ear a couple of seconds later, in a voice that keeps your pacing and tone. That is what Gemini 3.5 Live Translate does. Announced by Google on June 9, 2026, it is Google's audio-to-audio model for real-time, spoken translation across more than 70 languages. It shows up in three places today: the Google Translate app, Google Meet, and the Gemini Live API for developers. There is a free tier, and paid usage runs about $0.037 per minute of audio. If you have used the old "speak, wait, listen" translation flow, the shift here is that the conversation no longer has to stop.
What Gemini 3.5 Live Translate actually is
Google describes it as an audio-to-audio model that streams translated speech directly, rather than the transcribe-then-read-aloud flow of older tools. Those older tools converted your speech to text, translated the text, then ran text-to-speech, and every hop added a pause. Gemini 3.5 Live Translate processes audio as it streams and emits translated audio continuously, staying a few seconds behind you rather than waiting for you to finish a sentence.
Three things separate it from a basic translator. It automatically detects the language being spoken, so a single conversation can move across more than 70 languages and, in a group setting, 2,000-plus language combinations without anyone changing a setting. It preserves the speaker's intonation, pacing, and pitch, so the output sounds like a person rather than a robotic readout. And every piece of audio it generates is watermarked with SynthID, Google's tool for flagging AI-generated media. Per its DeepMind model card, it is built on the Gemini 3 Pro architecture, with a 131,072-token input window.
Where you can use it right now
Gemini 3.5 Live Translate is not a single product. It is one model surfaced in three different places, and how you "use" it depends on which one you mean.
In the Google Translate app
For everyday use, it is rolling out globally in the Google Translate apps on Android and iOS. You open Live translate mode, pick or let it detect the languages, and speak. On Android there is a new listening mode that plays the translation through the phone's earpiece, so a one-on-one conversation feels closer to wearing an interpreter's headset than passing a phone back and forth.
In Google Meet
Inside Google Meet, speech translation powered by this model is in private preview for enterprise customers. The jump is significant: Meet's earlier translation supported roughly five languages, and the new version pushes that to 70-plus languages and 2,000-plus language pairs in a single meeting. Each participant hears the discussion in their own language while people keep talking.
Through the Gemini Live API
For developers, the model is in public preview on the Gemini Live API and in Google AI Studio. This is the route if you are building translation into your own app, call center, or streaming product. Google has already lined up integrations with real-time audio platforms including Agora, LiveKit, Fishjam, Pipecat, and Vision Agents, and Grab is testing it for driver-passenger calls across a network that handles more than 10 million voice calls a month.
Calling the Live API
The model ID is gemini-3.5-live-translate-preview. You connect over the Live API's WebSocket (directly, or through the GenAI SDK for Python and JavaScript) and stream audio both ways. A minimal Python setup looks like this:
```python from google import genai from google.genai import types
client = genai.Client() config = types.LiveConnectConfig( response_modalities=["AUDIO"], translation_config=types.TranslationConfig( target_language_code="es-US", # BCP-47 target output_audio_transcription={}, # also return text ), )
async with client.aio.live.connect( model="gemini-3.5-live-translate-preview", config=config ) as session: await session.send_realtime_input(audio=mic_chunk) # 16 kHz PCM, ~100 ms async for msg in session.receive(): play(msg.data) # 24 kHz translated audio ```
The audio format is specific and worth getting right before you debug anything else. Input is raw 16-bit PCM at 16 kHz, mono, little-endian; output comes back as raw 16-bit PCM at 24 kHz, mono. Send your microphone audio in roughly 100 ms chunks. Translation behavior is set with a translationConfig block (full reference in Google's docs):
targetLanguageCode: the BCP-47 code you want output in (required).echoTargetLanguage: whether to repeat audio that is already in the target language.inputAudioTranscription/outputAudioTranscription: turn these on to also receive text transcripts of what went in and what came out.
Because this is a translation-focused model, a lot of the general Gemini toolkit is switched off: no function calling, no thinking, no structured outputs, no caching, no search grounding. It takes audio in and gives translated audio (plus optional transcripts) out, and nothing else. Sessions use ephemeral tokens with a 30-minute expiration window, which matters if you are handing credentials to a browser client.
Gemini 3.5 Live Translate pricing
There is a genuine free tier during the public preview: input and output are free of charge (subject to the usual preview rate limits), which is enough to prototype the whole flow before you commit.
On the paid tier, billing is per audio token at 25 tokens per second of audio:
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| Gemini 3.5 Live Translate | $3.50 (~$0.0053/min) | $21.00 (~$0.0315/min) |
| Gemini 3.1 Flash Live | $3.00 (~$0.005/min) | $12.00 (~$0.018/min) |
| Gemini 2.5 Flash Native Audio | $3.00 | $12.00 |
Combined, Live Translate works out to roughly $0.037 per minute of translated audio. Note this is per audio stream: a 30-minute two-way conversation bills both the incoming and outgoing streams, so budget on the order of a dollar or two rather than treating it as a flat per-call figure. The cheaper Gemini 3.1 Flash Live model exists, but it is a general live-audio model, not one tuned for the intonation-preserving, low-latency translation this one is built for. The price gap buys you translation quality, not just tokens. Figures are from Google's developer pricing page, checked July 15, 2026.
The limits worth knowing before you rely on it
Google's own model card is unusually candid here, and these are real enough to plan around rather than dismiss.
The synthesized voice can be inconsistent. It may shift after long pauses, change apparent gender, or get stuck on one voice during rapid multi-speaker exchanges, so a fast four-person meeting stresses it more than a calm two-person chat. Language detection is the other soft spot: it struggles with strong non-native accents, closely related languages, and rapid switching between languages mid-sentence. Background noise is filtered but not perfectly, and heavy noise can introduce artifacts in the translated audio. Finally, the API preview currently accepts audio input only, with no text-in path, and the whole thing carries preview status, which means behavior and availability can still change.
None of these are dealbreakers for a quiet one-on-one conversation. They are things to test before you put it in front of a noisy conference room or a call center.
How it stacks up against your other options
If you are choosing what to reach for, three questions sort it out:
- Casual, in-person, both people on phones? The Google Translate app version is free and needs no setup. It is the obvious pick.
- A multilingual meeting? Google Meet's enterprise translation is the path, once your organization has access to the preview.
- Building translation into your own product? The Live API is the only option that gives you programmatic control, and the pricing above is what you budget against.
Against the older turn-based approach (speak, wait for the beep, listen), the continuous model is a real step change for conversation flow. Against a human interpreter, it is far cheaper and instantly available in dozens of languages, but a professional still wins on nuance, idiom, and high-stakes accuracy, which the limitations above make clear.
If you are building this into a product, remember the speech layer and the text layer are separate bills: the translation comes from Google's Live API, while any summarizing or action-item extraction afterward runs on a text model you pick independently, where API pricing varies widely by provider.
Related reading
- Gemini Omni Flash pricing and API guide for Google's other real-time multimodal model.
- OpenAI Realtime API pricing if you are comparing per-minute costs across live voice providers.
FAQ
How do I activate live translate in Google Translate?
Open the Google Translate app on Android or iOS, choose the Live or conversation mode, and select your two languages or let the app auto-detect them. On Android, turn on listening mode to hear the translation through the earpiece.
How many languages does Gemini 3.5 Live Translate support?
It automatically detects more than 70 languages, and in a group setting it can handle over 2,000 language pairs within a single conversation.
Is Gemini 3.5 Live Translate free?
Yes, there is a free tier for both input and output. Paid usage on the Live API costs about $0.037 per minute of audio ($3.50 per 1M input tokens, $21.00 per 1M output tokens).
Can I use it on iPhone and Android?
Yes. The Google Translate app rollout covers both iOS and Android, though the earpiece listening mode is an Android feature at launch.
What happens when it detects the wrong language?
Because detection is automatic, strong accents, closely related languages, or fast language switching can cause a wrong guess mid-conversation, and the translated voice may also shift or change gender after long pauses. In a quiet one-on-one it rarely trips; in a noisy or fast multi-speaker setting, test it before you depend on it.
Choose the Translate app for personal conversations, Meet for multilingual teams, and the Live API when you need translation inside software you control. Whichever you pick, prototype on the free tier before you commit.