Google Ships Gemini 3.5 Live Translate Across 70 Languages to Developers
Google's Gemini 3.5 Live Translate hits the Gemini Live API in public preview, letting developers build real-time speech-to-speech translation across 70+ languages into their own apps.
Takeaways
- Public preview now live: Gemini 3.5 Live Translate is available in the Gemini Live API and Google AI Studio for developers to build with today.
- Audio-to-audio architecture: Collapses the traditional STT + MT + TTS pipeline into one model, translating across 70+ languages and 2,000+ language pairs with ~3-second latency.
- Prosody preservation: Translated speech retains the original speaker's intonation, pacing, and pitch, producing noticeably more natural output than generic TTS-based systems.
- Known limitations: Multi-speaker voice consistency is weak in rapid conversations; translation mode disables text input, system instructions, and tool use.
- $0.023/minute API pricing with partner integrations via LiveKit, Agora, Pipecat, and Fishjam; demo code available on GitHub.
- SynthID watermarking on all generated audio puts Google ahead of the EU AI Act's August 2026 synthetic content labeling deadline.
Real-time voice translation has been a demo-stage promise for years. Google just made it a developer primitive. Gemini 3.5 Live Translate is now available in the Gemini Live API in public preview, letting you wire continuous speech-to-speech translation directly into your own applications across more than 70 languages and 2,000+ language pair combinations.
One model, not a pipeline
The key architectural shift here is that Gemini 3.5 Live Translate is an audio model built on Gemini 3 Pro that streams speech-to-speech translation in over 70 languages, auto-detecting the source language and generating output that preserves the speaker's intonation, pacing, and pitch. That last part matters more than it sounds.
Traditional translation pipelines chain three separate models: a speech-to-text (STT) engine, a machine translation model, and a text-to-speech (TTS) synthesizer. Each hop adds latency and accumulates errors. Gemini 3.5 Live Translate collapses all three into a single audio-to-audio model. Unlike older systems that wait for a speaker to finish before translating, it processes speech as it flows, staying just a few seconds behind, and preserves intonation, tempo, and vocal pitch so the translated voice sounds closer to how the original speaker actually sounds.
On the technical side, the model accepts audio input with up to a 128K-token context window and produces audio and text output up to 64K tokens. The developer API ingests raw 16-bit PCM audio at 16 kHz mono, chunked at 100-millisecond intervals, and returns 24 kHz mono PCM output.
What the numbers actually say
Google describes the lag as "a few seconds behind the speaker." Independent benchmarking by LiveLingo Research, which measured the raw API endpoint across 120 test utterances and four language pairs, recorded a median first-audio latency of roughly 2,947 ms. That is a consistent ~3-second speaking delay, which matches Google's framing. Comprehension fidelity scored 4.93 out of 5 across English-to-Spanish, English-to-Chinese, English-to-Japanese, and English-to-German pairs, the strongest result among the competing systems tested.
What it is good at, and where it breaks
Google says the model handles background noise and overlapping voices, making it usable in real-world settings like busy cafes or open offices. Auto language detection requires no manual configuration, and it supports over 2,000 language combinations in a single meeting.
But Google's own model card is candid about the rough edges:
- Language detection struggles with non-native accents, closely related languages (think Portuguese vs. Spanish), or very rapid language switches mid-sentence.
- Multi-speaker voice consistency is a real limitation: voices can shift after long pauses, change gender, or get stuck on a single voice during rapid back-and-forth conversations with multiple participants.
- Late-resolving syntax in languages like Mandarin or Japanese (where the verb or polarity comes at the end of a sentence) can cause the model to commit to a translation before it has enough context, occasionally inverting the meaning.
- Translation mode is constrained: text input is not supported, and the broader Gemini tool ecosystem (system instructions, function calling) is unavailable in this mode.
How to build with it
The developer entry point is the Gemini Live API, which streams speech in and out so you get continuous behavior rather than batch translation.
The model ID is gemini-3.5-live-translate-preview, accessible today through the Google AI Studio playground or directly via the API.
A minimal Python session looks like this:
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
)
)
async with client.aio.live.connect(model=model, config=config) as session:
async for response in session.receive():
# handle translated audio chunks here
passBecause the hardest part of any real-time voice app is the media transport layer (WebRTC, jitter buffers, echo cancellation), Google leans on partner platforms rather than asking you to own that infrastructure. Developer platforms like Agora, Fishjam, LiveKit, Pipecat, and Vision Agents integrate the Gemini Live API, handling the complex real-time media streaming infrastructure so developers can focus on the user experience. Google has also published a LiveKit demo on GitHub showing a broadcast pattern where a single Gemini session per target language serves all attendees who selected that language, avoiding redundant API calls at scale.
The broadcast architecture worth stealing
The LiveKit demo illustrates a clean scaling pattern for multi-language broadcasts:
- The host publishes audio into a LiveKit room.
- A translation bridge subscribes to that audio track.
- One Gemini Live API session is created per target language (not per listener).
- All attendees who selected Spanish share the same translated audio stream.
This session-sharing approach means fifty Spanish listeners do not spin up fifty identical Gemini sessions. It is the kind of infrastructure detail that separates a demo from a production deployment.
Pricing and availability
API pricing is $0.023 per minute, which undercuts several competing services. The consumer Google Translate app is free and already rolling out globally on Android and iOS. The 70+ language expansion enters private preview for enterprise customers in Google Meet. General availability timing, pricing, and Workspace tier requirements for Meet have not been announced.
Every audio output is watermarked using Google's SynthID technology, embedding an invisible marker in the audio to flag it as AI-generated, which matters because the EU AI Act's Article 50 requiring synthetic content labeling comes into force in August 2026.
The bigger picture
In December 2025, Google brought Gemini-powered text translation to Search and the Translate app. In February 2026, it added contextual and tone-aware features. This release extends live speech translation to consumers globally, enterprise customers in preview, and now to developers as a buildable primitive. The pattern is deliberate: Google is positioning translation as a platform layer that other products draw from, not just a feature it ships in its own apps.
Grab is testing the model to enable multilingual communication in near real-time between drivers and travelers at pickups, and those users make over 10 million voice calls per month through the platform. That is the kind of production signal that tells you this is past the demo stage. The question for builders is no longer whether real-time voice translation is technically feasible at scale. It is whether the 3-second lag, the multi-speaker voice drift, and the translation-mode API constraints fit the specific product you are trying to ship.