Google has unveiled Gemini 3.5 Live Translate, a new speech‑to‑speech model designed to deliver continuous, near real‑time translation across more than 70 languages. The update represents one of the company’s most significant steps toward low‑latency multilingual communication since the launch of Google Translate, which now processes over a trillion words each month.
Unlike earlier turn‑based systems, the new model simultaneously produces translated speech as the speaker talks, typically staying only a few seconds behind. Google says the system preserves intonation, pacing, and pitch to generate more natural‑sounding output. It also automatically detects languages without requiring manual input and is engineered to perform reliably within noisy or unpredictable environments.
A major component of the rollout is its integration into Google Meet, where the company is expanding translation support from a limited set of English‑centric pairs to more than 2,000 language combinations. Previously, Meet’s live translation features were restricted to only to and from English workflows. The new model allows supported languages to be translated directly into any other, significantly broadening access for global teams, classrooms, and multilingual events.
Developers can begin testing the model through the Gemini Live API and Google AI Studio, now available in public preview. Enterprise users will see the Meet integration in private preview this month, with a wider release planned later in the year for general users on Android and IOS. The Google Translate mobile app is also receiving updates, including a new Android “listening mode” that routes translated audio through the phone’s earpiece for more discreet use.
Early pilot partners including Grab, CJ ENM, LiveKit, Vision Agents, and Agora reported improvements in latency, accuracy, and automatic language detection. Grab, for instance, is evaluating the system to support multilingual communication across its 10 million monthly voice calls.
All audio generated by the model is said to be watermarked using SynthID, Google’s imperceptible watermarking technology intended to help identify AI‑generated content and mitigate misinformation risks. For additional details on its approach to safety and responsibility, please review the model card.
