Google's Gemini 3.5 Live Translate Streams 70+ Languages Without Waiting

Real-time voice translation has been a hard problem for a long time. Most systems work in turns: wait for the speaker to finish, transcribe, translate, synthesize speech. The result is choppy, robotic, and conversationally awkward. Google just shipped something fundamentally different. Gemini 3.5 Live Translate is a new audio-to-audio model that streams translated speech while the source speaker is still talking, staying just a few seconds behind, and doing it across more than 70 languages.

Not turn-by-turn. Continuous.

Unlike turn-by-turn systems that wait for the speaker to finish speaking before responding, 3.5 Live Translate generates speech continuously, balancing the trade-off between waiting for context to improve quality and translating immediately to stay in sync with the speaker. That tradeoff is the crux of the whole engineering challenge: translate too eagerly and you get garbage because you haven't heard the full sentence; wait too long and the conversation feels like a bad satellite call.

Gemini 3.5 processes streamed audio continuously while retaining enough context to interpret the developing sentence, then generates translated speech with limited delay, balancing immediate output against the need to understand what the speaker means. The model makes split-second decisions about when it has heard enough to commit to a translation.

Beyond just getting the words right, the model automatically detects 70+ languages and generates smooth, natural-sounding translated speech that preserves the speakers' intonation, pacing and pitch. Traditional speech translation often turns spoken language into text, translates the text, and reads the result through a generic text-to-speech voice , a process that can communicate meaning while losing the pauses, emphasis, rhythm, and emotion. Gemini 3.5 Live Translate skips the intermediate text step entirely.

Under the hood

Gemini is a natively multimodal model, meaning it was trained on text, images, code, and audio simultaneously rather than being a language model with audio features bolted on. For translation specifically, this matters because speech carries emotional and prosodic information that text simply doesn't encode.

According to the official model card, Gemini 3.5 Live Translate is built on top of Gemini 3 Pro, with a 128K token input context window and up to 64K token output. The model was evaluated across three dimensions: translation quality (using AutoMQM, an error-based metric that categorizes translation mistakes), latency (measured both at the word level and at session start), and speech naturalness (checking for choppy audio, voice drift, and artifacts).

On the developer side, each session uses raw 16-bit PCM audio at 16 kHz mono as input and produces 24 kHz mono PCM audio as output, sent in 100-millisecond chunks. All generated audio carries Google's SynthID watermark , an imperceptible signature woven into the waveform that allows downstream systems to identify the audio as machine-generated. This matters for compliance: the EU AI Act's Article 50, requiring synthetic content labeling, comes into force on August 2, 2026, and SynthID watermarking puts Google ahead of that deadline.

Where it falls short

Google's own model card is refreshingly candid about the limitations. Voices can be inconsistent, and voices may shift after long pauses, change gender, or get stuck on one voice during rapid multi-speaker sessions. Language detection can struggle with non-native accents, similar languages, or rapid language switches.

Quality also isn't uniform across all 70+ languages. Translation between closely related languages (like Spanish and Portuguese) tends to be more accurate than translation between languages with very different structures (like English and Japanese). Live Translate is designed for comprehension, not verbatim accuracy. That's an important distinction if you're building anything where precision matters , legal, medical, or compliance contexts will need human review.

Where to use it

The model is rolling out across three surfaces simultaneously:

• Consumer: Consumers get the update immediately in the Google Translate app on Android and iOS. When used with earbuds, the translated audio plays back preserving the speaker's tone. Android users also get a new Listening Mode: hold the phone to your ear like a normal call and hear a private translation through the speaker , no earbuds needed.
• Enterprise: The expansion of Google Meet from 5 languages to 70+, and from English-only pairs to more than 2,000 language combinations, is the most operationally significant change for enterprise users. Google Meet is entering private preview for select Workspace customers now, with a broader rollout later this year.
• Developers: Developers can access the model in public preview via the Gemini Live API and Google AI Studio. API pricing is $0.023 per minute, which undercuts several competing services.

For developers who don't want to handle raw WebSocket audio streaming themselves, developer platforms like Agora, Fishjam, LiveKit, Pipecat, and Vision Agents enable developers to build and deploy voice translation apps with ease, handling the complex real-time media streaming infrastructure so developers can focus on the user experience.

Real-world deployment: Grab's 10M monthly calls

The most telling signal of production readiness is who's already building on it. Grab is testing Gemini 3.5 Live Translate to enable multilingual communication in near real-time between drivers and travelers at pickups , a user base that makes more than 10 million voice calls per month. The driver-traveler pickup scenario involves two parties who may not share a language trying to coordinate a specific physical location, a scenario where miscommunication has direct operational consequences. Early feedback from partners highlights impressive translation quality, accuracy, and low latency.

What this changes

The real shift here isn't just better translation quality. It's the architectural move from a pipeline (speech → text → translate → synthesize) to a single end-to-end streaming model. That collapse removes latency at every step, and it's what makes the "a few seconds behind" lag possible at all.

The technical shift is significant: translation is becoming a live conversational interface that must manage language, context, timing, noise, speaker changes, and vocal expression together. In 2016, Google switched to neural translation, dramatically improving the quality of written translations. Today, the focus shifts from simple text conversion to the reproduction of a lively conversation, with voice, rhythm, and emotional nuance , it's no longer just about understanding words, but about preserving the "how" they're said.

Google isn't the only company moving in this direction: Microsoft, Meta, and specialized startups are also investing in real-time voice translation and multimodal models. But shipping a model at this quality level across consumer apps, an enterprise meeting platform, and a public API simultaneously is a meaningful lead. The practical use cases are immediate: multilingual customer support, live conference interpretation, cross-border sales calls, language learning, and real-time dubbing for content creators. A sales call, a support conversation, or an internal all-hands meeting that previously required a human interpreter or was restricted to a small set of supported languages can now proceed in any of the 70+ supported languages without additional infrastructure.