Skip to main content
Back

Google Ships Gemini 3.5 Live Translate Across 70 Languages in Near Real-Time

Google's Gemini 3.5 Live Translate delivers near real-time speech-to-speech translation across 70+ languages while preserving tone, pace, and pitch

  • New model: Google released Gemini 3.5 Live Translate, a dedicated speech-to-speech translation model supporting 70+ languages.
  • Continuous streaming: Unlike turn-based systems, it translates audio as it streams, staying just seconds behind the speaker with no awkward pauses.
  • Voice preservation: The model preserves the speaker's intonation, pacing, and pitch in the translated output.
  • Three surfaces: Available now in Google Translate (Android/iOS), via the Gemini Live API in public preview, and in private preview for Google Meet enterprise customers.
  • Google Meet upgrade: Expands Meet's translation from 5 languages (English-only pairs) to 70+ languages and 2,000+ language combinations.
  • Real-world test: Grab is piloting it for driver-traveler calls, a service handling over 10 million voice calls per month.

Google just shipped Gemini 3.5 Live Translate, a dedicated audio model that converts live speech into over 70 languages in near real-time. It is available right now in the Google Translate app on Android and iOS, in public preview via the Gemini Live API and Google AI Studio, and rolling out in private preview for enterprise Google Meet customers.

This is not a wrapper around a general-purpose model. Google built 3.5 Live Translate as a purpose-built audio pipeline, separate from the conversational Gemini Live agents, optimized for one thing: getting translated speech out as fast and naturally as possible.

The problem with every other translator

Traditional speech translation systems work in three sequential steps: transcribe the audio, translate the text, then synthesize new speech. Each step adds latency, and the whole pipeline waits for the speaker to finish a sentence before doing anything. The result is the familiar awkward pause-and-burst pattern that makes real conversations feel robotic.

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. The model stays just a few seconds behind the speaker throughout the session, rather than accumulating a growing lag.

The other big problem was voice identity. Most translation systems produce a generic synthetic voice regardless of who is speaking. The model automatically detects 70+ languages and generates smooth, natural-sounding translated speech that preserves the speakers' intonation, pacing and pitch. That means a speaker's excitement, hesitation, or emphasis carries through into the translated output.

How it actually works

Under the hood, the model operates as a streaming audio-to-audio pipeline. The Gemini Live API supports low-latency, real-time speech-to-speech translation between 70+ languages using the gemini-3.5-live-translate-preview model.

The API design reflects the specialized nature of the task. Here is what it looks like to spin up a translation session in Python:

from google import genai
from google.genai import types
client = genai.Client()
model = "gemini-3.5-live-translate-preview"
config = types.LiveConnectConfig(
    response_modalities=["AUDIO"],
    input_audio_transcription=types.AudioTranscriptionConfig(),
    output_audio_transcription=types.AudioTranscriptionConfig(),
    translation_config=types.TranslationConfig(
        target_language_code="es",  # translate to Spanish
        echo_target_language=True   # echo back if speaker already speaks Spanish
    )
)
async with client.aio.live.connect(model=model, config=config) as session:
    # stream 16-bit PCM audio at 16kHz in 100ms chunks
    # receive translated audio at 24kHz

The audio format requirements are strict: input must be raw 16-bit PCM at 16kHz, and output comes back at 24kHz. Chunks should be sent in 100ms windows to hit the latency targets. Crucially, only audio input is supported for translation. Text input is not supported. This constraint is intentional: it keeps the pipeline lean and latency predictable.

Live Agent vs. Live Translation: two different mental models

It is worth being precise about what this model is and is not. The Gemini Live API already powers conversational agents, but 3.5 Live Translate is a fundamentally different mode. The key differences:

  • No tool use. Unlike Live agents, the translation model has no support for function calling, Google Search, or system instructions. It is a pure translation pipeline.
  • Continuous stream, not turns. Agents rely on pause detection and intent recognition. The translator processes audio as a continuous stream without waiting for turn boundaries.
  • Audio-only input. Agents can handle text, images, and video. The translator accepts only audio, which is what keeps latency tight.
  • Simpler config. You set a target_language_code and a boolean echo_target_language flag. That is essentially the entire configuration surface.

The echo_target_language flag is a practical detail worth understanding: when set to true, if the speaker is already talking in the target language, the model echoes that audio back rather than going silent. This matters for multi-speaker scenarios where participants switch languages mid-conversation.

Where it ships

The rollout covers three surfaces simultaneously:

  • Google Translate app (Android and iOS): Available globally now. Connect headphones and the translated audio streams to your ears. For Android users, Google is also starting to roll out a new listening mode with 3.5 Live Translate that lets you hear translations directly through your phone's earpiece. Hold the phone to your ear like a call, and translated audio comes through without others hearing it.
  • Gemini Live API / Google AI Studio: In public preview for developers now. The model ID is gemini-3.5-live-translate-preview.
  • Google Meet: Launching in private preview for select business Google Workspace customers starting this month, followed by a broader rollout later this year. The upgrade is significant: Meet's previous speech translation supported only five languages and could only translate to and from English. The new version supports 70+ languages and over 2,000 language pair combinations in a single meeting.

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

Partners at Grab are testing the model to enable multilingual communication in near real-time between drivers and travelers at pickups. These users make over 10 million voice calls per month through Grab. That scale makes it a meaningful stress test for latency and noise robustness in unpredictable real-world environments.

The model handles multilingual inputs without the need to manually configure settings. At the same time, its noise robustness ensures applications can handle loud, unpredictable environments. For developers building on top of it, platforms like LiveKit, Pipecat, Agora, and Fishjam already have integrations that handle the real-time media streaming infrastructure.

Known limitations

The official docs are refreshingly honest about where the model struggles:

  • Voice replication inconsistency. Voices might shift after long pauses, assign the wrong gender based on how the speech starts, or get stuck on one voice during rapid multi-speaker conversations.
  • Language detection edge cases. Language detection struggles with heavy accents, similar languages (e.g., Spanish vs. Portuguese), or rapid language switches. The final translation should still be accurate, but the input transcript language code may be wrong.
  • Background audio. The model filters noise and music, but not perfectly. When echoTargetLanguage is enabled, background noise can introduce artifacts.

Pricing and access

The model is available in public preview via the Gemini API. Preview models may change before becoming stable and have more restrictive rate limits. Specific per-token pricing for the Live Translate model has not been published separately from the Live API billing structure, which charges per turn based on the session context window. Google AI Studio provides a free sandbox for testing. For production deployments, you will need a paid API key and should expect Live API billing to apply.

All audio generated by the model is watermarked using SynthID, Google's imperceptible audio watermarking system, so AI-generated translations remain detectable even after processing. For teams building multilingual communication tools, live interpretation services, or cross-language call infrastructure, this is the most accessible production-ready API for the task that currently exists.