# Gemini 3.5 Live Translate entered public preview for developers > Google released a streaming speech-to-speech model for more than 70 languages through the Gemini Live API and Google AI Studio. Canonical URL: https://www.devobs.io/news/news-gemini-3-5-live-translate-release/ By: Amara Okafor Published: 2026-09-06T11:58:54.627Z Updated: 2026-09-06T11:58:54.627Z Event date: 2026-06-09 Section: AI Google released Gemini 3.5 Live Translate, a streaming speech-to-speech model that automatically detects more than 70 languages. The [June 9 announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-live-3-5-translate/) said the model generates translated speech continuously while balancing additional context against delay. Developer access began in public preview through the Gemini Live API and Google AI Studio, alongside rollouts in Google Translate and a private preview for selected Google Meet customers. ## Streaming translation is a latency policy A live system cannot wait indefinitely for a complete sentence. It must decide when enough context exists to emit speech, then cope with corrections, interruptions and language switches. That makes partial-result semantics as important as aggregate translation quality. A client needs to know whether emitted audio is final, revisable or replaceable after more context arrives. Google reported language coverage and product availability, but applications still need their own evaluations for accents, domain vocabulary, noisy rooms and overlapping speakers. Safety-critical or contractual conversations require a fallback because plausible speech can conceal a semantic error. ## Test the conversation, not isolated clips A useful prototype should measure end-to-end lag, interruption recovery and speaker-turn handling over full sessions. Include code-switching, names, numbers, negation, poor connectivity and a participant who corrects themselves. Make the current source and target languages visible even when detection is automatic. Teams should also define consent, retention and redaction for streamed audio and transcripts. Provide a way to pause translation and recover the original utterance. The preview lowers the barrier to multilingual voice features; dependable deployment still requires domain-specific evaluation and clear failure behavior. ## Source references -