Our Take
The latency claim is real but understated: staying 'a few seconds behind' is the hard part of simultaneous translation, and Google is shipping it across three products at once.
Why it matters
Simultaneous voice translation at scale has been the missing piece in real-time multilingual meetings and customer support. Grab's 10 million monthly driver-rider calls signal enterprise demand.
Do this week
Platform leads: Test 3.5 Live Translate integration via Gemini Live API this week so your roadmap can move real-time translation from roadmap to Q3 feature.
Gemini 3.5 Live Translate ships across Google products
Google released Gemini 3.5 Live Translate, an audio model for simultaneous speech-to-speech translation across 70+ languages. The model runs continuously during conversation, balancing latency against translation quality to keep output within a few seconds of the speaker.
The rollout spans three channels: developer access via Gemini Live API and Google AI Studio (public preview); Google Meet for select Google Workspace customers (private preview, expanding later this year); and the Google Translate app on Android and iOS (global rollout). Key capabilities include automatic language detection, preservation of speaker intonation and pacing, and noise robustness for unpredictable environments.
Google Meet gains the largest functional expansion. The previous version supported only five languages and translated exclusively to and from English. The new version enables 2,000+ language-pair combinations in a single meeting, a structural change for multinational enterprises.
Early adopters already testing the model include Grab (ride-hailing, 10 million monthly voice calls), Agora, Fishjam, LiveKit, Pipecat, and Vision Agents (all platform providers integrating the Gemini Live API). Grab's test case focuses on real-time communication between drivers and passengers at pickups.
All audio output is watermarked with SynthID, an imperceptible marker embedded to identify AI-generated speech and reduce misuse.
Simultaneous translation removes a critical meeting friction
Until now, live translation either waited for speakers to finish sentences (turn-based) or required manual language configuration. Both break meeting flow. A few seconds of latency is the practical ceiling for fluent conversation; Gemini 3.5 aims to stay within it while handling noise, multilingual input, and 70 language combinations on a single call.
The deployment across consumer (Google Translate), workplace (Google Meet), and developer APIs in parallel suggests Google sees three distinct markets: casual travel translation, enterprise meetings, and embedded voice products. Grab's use case points to a fourth: customer support at scale in regions with mixed-language populations.
The expansion of Meet from five to 70+ languages is not a cosmetic feature. It unlocks translation workflows that were impossible before (e.g., Japanese, Spanish, and Portuguese speakers in one call without English as a bridge language). For companies with distributed teams across multiple continents, this removes a technical blocker to async and real-time collaboration.
Integrate and benchmark latency in your deployment path
If your product involves voice calls, customer support, or meetings across language groups, Gemini Live API access is live. Test latency and translation quality against your actual audio (noise profiles, speaker accents, domain jargon) before committing to a product roadmap. The "few seconds behind" claim is generic; your tolerance for lag is specific.
For Google Workspace administrators, flag that Meet translation is rolling out in phases. Plan comms for your user base now, especially if teams use five-language workarounds that will become obsolete.