Gemini 3.5 Live Translate: Real-Time Speech Translation Across 70+ Languages

Gemini 3.5 Live Translate is Google’s specialized model for streaming spoken language into translated speech. Released on June 9, 2026, it supports more than 70 languages and is designed to stay only a few seconds behind the speaker rather than waiting for a complete turn.

That continuous behavior is the defining feature. Conventional speech translation often follows a chain—transcribe, wait for a pause, translate, synthesize—and creates awkward silence. Live Translate balances two competing needs: wait long enough to understand context, but speak soon enough to remain useful in a live conversation.

The model is available through the Gemini Live API under the preview identifier gemini-3.5-live-translate-preview. It is also rolling into Google products, but product availability differs: Google Translate received a broader app rollout, while the Google Meet upgrade began as a private preview for selected business customers.

A translator, not a general voice agent

Google explicitly distinguishes Live Translation from a Live Agent. A Live Agent participates in turn-based conversation, reasons, follows instructions, and may call tools. Live Translate behaves like an interpreter: it continuously processes incoming audio and produces translated audio.

This narrower scope is a strength for latency. It is also a boundary developers must respect. The translation model cannot look up a technical term on the web, query a customer record, or decide to send a follow-up email.

How continuous translation works for users

The developer sends a live audio stream and configures a target language using its BCP-47 code. The service returns translated speech as audio. Optional input and output transcription settings can provide text transcripts alongside the stream.

An echo-target-language setting controls what happens when the incoming speech already matches the target. The model can repeat it or remain silent. This is useful in multilingual rooms where participants switch languages, but the choice should be visible to users; otherwise silence may be mistaken for a broken connection.

Google says the model automatically detects languages and preserves aspects of the speaker’s intonation, pacing, and pitch. The result aims to sound like one continuous interpretation rather than a sequence of robotic sentences.

Language coverage and product rollout

More than 70 languages create over 2,000 possible source-target combinations. Google’s launch examples include multilingual meetings, travel, lessons, broadcasts, and communication between marketplace users.

Google Translate began receiving the model globally on Android and iOS, including a headphone experience and an Android listening mode. Google Meet’s planned upgrade expands speech translation from five languages to more than 70 and removes the earlier requirement to translate only to or from English, but that Meet rollout began in private preview.

These product releases should not be used as proof that every API language pair performs equally well. Dialects, names, technical terminology, code-switching, and noisy environments vary substantially. A production team needs an evaluation set from its own users.

Model limits are unusually explicit

The Live API guide documents several important limitations:

• only audio input is supported in translation mode;
• voice replication can shift after a long pause;
• the wrong gender may be inferred from the beginning of speech;
• rapid multi-speaker conversation may collapse into one voice;
• heavy accents and closely related languages can confuse detection;
• rapid language switching is difficult;
• background noise or music may not always be ignored;
• echoing target-language input can introduce audio artifacts.

The model page also lists no support for function calling, code execution, file search, search grounding, structured output, thinking, caching, Batch API, or priority inference. Those omissions confirm that this is a dedicated real-time channel, not a general Gemini inference endpoint.

Latency is an application property too

Google describes near-real-time behavior, but end-to-end delay includes capture, encoding, network transit, service processing, playback buffering, and the receiving device. A model that stays a few seconds behind can still feel slow if an application adds several seconds of buffering.

Measure:

• time from spoken phoneme to translated playback;
• interruption and overlap behavior;
• missing or duplicated audio;
• reconnect time after a network drop;
• transcript and audio alignment;
• performance on low-bandwidth mobile networks.

The launch post highlights integrations with real-time platforms such as LiveKit, Agora, Fishjam, and Pipecat. Those systems solve parts of the media transport problem; the application still owns session state and user experience.

Translation quality needs human-centered tests

Word error rate alone cannot capture whether live translation works. Evaluate semantic accuracy, names and numbers, politeness, emotion, pacing, and the timing of corrections. Include speakers with different accents and microphones.

For regulated or high-stakes contexts, live machine translation should not replace a qualified interpreter. A small error in medical, legal, safety, or financial speech can be consequential even if the conversation sounds fluent.

Provide users with a clear way to pause translation, view transcripts, correct names, and report errors. If the system is uncertain or loses the stream, it should signal that state rather than fill silence with plausible speech.

Test conversational repair as well as first-pass accuracy. Speakers often correct themselves, interrupt, or finish each other’s sentences. A useful system must decide whether to revise an earlier translation, continue with a clarification, or remain silent. The interface should make revisions understandable instead of playing contradictory audio with no explanation. For meetings, preserve timestamps so a reviewer can align source speech, translated speech, and both transcripts when investigating a complaint.

Cost and operational planning

Google’s model page does not publish a simple flat session price next to the capability table. Consult the current Gemini API pricing for the exact audio rates and region. Beyond model usage, budget for streaming infrastructure, relay servers, recording or transcript storage, observability, and support.

A useful unit is cost per translated participant-minute, including failed and idle sessions. Disconnect inactive streams, set session limits, and avoid keeping a paid channel open when no speech is present.

Safety, consent, and SynthID

Google says audio generated by its models is watermarked with SynthID. Watermarking supports detection but does not remove the need to disclose that listeners hear synthesized translation.

Participants should know which audio is captured, which language is generated, whether transcripts are stored, and who can access them. Voice preservation is not the same as an authorized voice clone. Do not market the output as an exact reproduction of a speaker’s identity.

From live interpretation to a polished localized video

Live Translate is ideal for a meeting or call; a publishable localized video has different requirements. It needs a stable script, repeatable voice, editorial review, timing control, and mouth synchronization. Those are offline production tasks, not features of the translation-only Live API.

A team could use Live Translate during an interview, preserve an approved transcript, and then produce a reviewed localized deliverable through separate tools. With consent, Codex or Claude can call Medux to synthesize the final narration, adjust pitch where appropriate, and synchronize it to an avatar. Medux is not a native Gemini integration, and the workflow does not turn Live Translate into a voice-cloning model.

Relevant tutorials show the distinct steps: text-to-speech and authorized voice cloning, audio pitch adjustment, and lip synchronization.

Keep the live session and production job separate:

1. obtain participant consent and capture the approved transcript;
2. have a fluent reviewer edit the translation;
3. approve voice and avatar rights;
4. create one offline TTS job;
5. review pacing before any pitch adjustment;
6. create the lip-sync task and validate the output;
7. disclose the synthesized localization when published.

Live Translate removes conversational delay across languages. The offline pipeline turns reviewed language into a durable media asset. Combining them is useful precisely because each layer keeps a clear role.