Traditional speech translation has an obvious turn-taking rhythm: wait for the speaker to finish, transcribe the sentence, translate it, then synthesize a new voice. The result may be accurate, but the silence between turns makes conversation feel unnatural.
Google’s Gemini 3.5 Live Translate uses continuous streaming instead. It receives speech, detects the language and context, and produces translated audio while the speaker is still talking. The model supports more than 70 languages and is available through the Gemini Live API, with the same technology feeding live translation experiences in Google Translate and Google Meet.
It is not a general-purpose Live Agent
Both products use the Live API, but their operating models are different.
A regular Live Agent is an assistant. It detects turns and intent, handles interruptions, and may call tools. Live Translate is a specialized interpretation pipeline: its input and output are audio, and it does not support tools, instructions, or text input.
That restriction is deliberate. A simultaneous translation system constantly balances two questions: does it have enough context to translate correctly, and should it begin speaking now? Waiting produces more context but adds latency; speaking too early risks committing to a meaning that later words may change.
“Real time” therefore does not mean zero delay. It means translated speech follows the speaker continuously rather than restarting a full processing cycle after every sentence.
Three capabilities that matter
Automatic source-language detection
An API session only needs a target language. The model detects whether the incoming speech is Chinese, Japanese, Spanish, or another supported language and translates it to the configured output.
This is useful for travel and multilingual meetings, but detection has limits. Google notes that heavy accents, closely related languages, and rapid language switching can cause problems.
Preserving how someone speaks
The output is not simply translated text read by a fixed synthetic voice. The model attempts to retain the original speaker’s intonation, pacing, and pitch so that emphasis and rhythm survive the translation.
It should not be confused with exact voice cloning. Voice characteristics can shift after long pauses, and rapid multi-speaker conversations may assign the wrong voice. “Preserving delivery” is a more accurate description than “replicating a voice.”
Filtering some background audio
The model attempts to suppress noise and music when producing clean translated speech, which matters in streets, restaurants, and transit hubs. The documentation also makes clear that not every background sound will be filtered consistently.
Minimal Python configuration
The preview model is named:
gemini-3.5-live-translate-previewThe core session configuration looks like this:
from google.genai import types
config = types.LiveConnectConfig(
response_modalities=["AUDIO"],
input_audio_transcription=types.AudioTranscriptionConfig(),
output_audio_transcription=types.AudioTranscriptionConfig(),
translation_config=types.TranslationConfig(
target_language_code="en",
echo_target_language=True,
),
)target_language_code uses BCP-47 identifiers. Simplified Chinese is zh-Hans, Traditional Chinese is zh-Hant, Japanese is ja, and English is en.
echo_target_language controls what happens when the input is already in the target language. When set to true, the audio is echoed; when false, the model remains silent. The default is false.
The transcription options are optional. Enabling them returns text for both the input speech and translated output, which is useful for bilingual captions or conversation records.
Audio requirements
Live Translate expects specific formats:
- Input: raw 16-bit PCM, 16 kHz, mono, little-endian;
- Output: raw 16-bit PCM, 24 kHz, mono, little-endian;
- Recommended chunk duration: 100 milliseconds;
- Translation mode accepts audio input only, not text.
Model access is only one part of a production integration. Applications still need microphone capture, resampling, WebSocket management, playback buffering, reconnect behavior, and echo cancellation. Integrations with real-time media platforms such as Agora, LiveKit, and Pipecat can reduce the amount of low-level audio infrastructure a team must build itself.
Do not expose API keys in client applications
A browser or mobile client connecting directly to the Live API should not contain a long-lived API key. Google provides ephemeral tokens: a server creates a short-lived, limited-use credential that the client uses to establish the session.
The default recommendation is to lock translation settings when issuing the token so the client cannot alter important fields. If users need to select their own target language, the server can explicitly leave that configuration unlocked.
Ephemeral tokens currently use the v1alpha endpoint. Even for a preview model, the authentication boundary should be designed before a product ships, not added after a key has already reached client code.
Where it fits
Likely applications include:
- Multilingual video meetings and customer-support calls;
- Face-to-face conversations while traveling;
- Voice calls between drivers and passengers;
- Personal interpretation through headphones at international events;
- Live products that need translated audio and bilingual captions.
It does not replace every translation workflow. Legal documents, medical instructions, and contracts still require full context, controlled terminology, and human review. Products also need fallback behavior for overlapping speakers, heavy accents, or complicated background audio.
From translating sentences to translating conversation
The important change is not just the number of supported languages. Earlier systems mostly translated an utterance after it was complete. Gemini 3.5 Live Translate participates in the conversation as it unfolds, continuously trading off latency, context, and vocal expression.
For users, that moves phones and headphones closer to personal interpretation devices. For developers, it offers a translation pipeline that can be embedded in calls, meetings, and travel products. The model is still in Preview, however, and its known limitations around voice consistency, detection, and background audio should be tested with realistic input before launch.
