MOUNTAIN VIEW, Calif.: Google DeepMind launches Gemini 3.5 Live Translate, a new audio model that delivers real-time speech-to-speech translation across more than 70 languages.
The model automatically detects which language someone is speaking without manual setup. Translated speech preserves the original speaker’s intonation, pacing, and pitch. The system supports more than 2,000 language pairings at launch globally.
✨ AI Summary
Google DeepMind has launched Gemini 3.5 Live Translate, an audio model providing real-time speech-to-speech translation across 70 languages. The system continuously streams audio while preserving the speaker’s original intonation and pitch. It features automatic language detection and operates with low latency, even in noisy environments.
The model is currently rolling out across Google Translate, Google Meet, and developer platforms. All generated audio includes SynthID watermarking to ensure transparency and address deepfake concerns. This release expands Google’s translation capabilities for enterprise and consumer users, providing accessibility through various hardware and software applications.
Traditional translation tools wait for a complete sentence before generating output to users. Gemini 3.5 Live Translate streams translated speech continuously while the speaker is talking. The model stays only a few seconds behind the original audio throughout sessions.
The architecture handles noisy environments without degraded performance under stress conditions. Multilingual inputs work without manual switching between language tracks during conversations. Auto-detection eliminates the friction of selecting languages before each interaction begins.
The rollout spans multiple Google products immediately upon today’s announcement. Google Translate adds the model on Android and iOS apps simultaneously. The Gemini Live API and Google AI Studio open public preview access for developers globally.
Google Meet receives the model in private preview for enterprise customers this month. Android phones gain a new listening mode that delivers translations through the device earpiece. The hardware-agnostic approach means no headphones or dedicated devices are necessary.
SynthID watermarking marks all AI-generated audio output from the system automatically. The move addresses growing concerns around deepfake voice content across consumer platforms. Google positions the watermarking as essential transparency for live translation use cases.
“Our latest audio model takes real-time speech translation to the next level by delivering low-latency translation,” says Google in the launch announcement.
The launch intensifies pressure on Microsoft Translator, DeepL, and Meta’s translation efforts across enterprise and consumer markets. Google previously dominated translation research for over two decades.
Previous
