Google drops near-instant voice translation that keeps your tone

The most awkward part of cross-language calls is that painful three-second pause while an app transcribes, translates, and spits out robotic audio.

Google is trying to make that friction disappear. The company just released Gemini 3.5 Live Translate, a dedicated audio model designed for continuous, near real-time speech-to-speech translation across 70+ languages. Instead of waiting for a turn to end, it translates on the fly just seconds behind the speaker while keeping their original pitch, tone, and pacing intact.

Real-time streaming over turn-taking

Traditional speech translation works like a walkie-talkie: you speak, stop, and wait. Gemini 3.5 Live Translate switches entirely to continuous audio streaming, dynamically balancing context gathering with immediate delivery to prevent awkward pauses.

To achieve strict low latency, Google stripped away the bloat. According to developer documentation, unlike full conversational Gemini agents, this translation model drops function calling, web search, system instructions, and video inputs. It strictly takes audio in and pushes translated audio out.

Every audio stream generated by the model includes Google's imperceptible SynthID watermark to flag AI-generated speech.

Ear-to-phone listening and multi-language Meet calls

Developers can test the model today in Google AI Studio and via the Gemini Live API, with real-time platforms like Agora, LiveKit, and Pipecat handling the streaming infrastructure. Grab is already testing it to bridge calls between drivers and travelers, a system handling over 10 million calls monthly.

For consumers, it is rolling out globally on Google Translate for Android and iOS. On Android, a new listening mode lets you hold the phone to your ear like a standard phone call, streaming private live translation of a museum tour or lecture without headphones.

Google Meet is also getting an overhaul in private preview for select Workspace enterprise customers. Meet previously handled only 5 languages and required translating to or from English. With the new model, it supports 70+ languages and over 2,000 language combinations in a single meeting.

Why it matters

For remote teams, international travelers, and cross-border operators, this is an immediate quality-of-life jump. Holding a phone to your ear like an ordinary call to get private simultaneous translation solves a very real, everyday travel annoyance without requiring specialized earbuds.

Strategically, Google made a smart product decision here. Instead of trying to force a bloated, multimodal reasoning agent into a real-time call, it carved out a stripped-down, audio-only pipeline optimized purely for speed and acoustic fidelity. It pushes the rivalry with dedicated hardware translation gadgets and standalone tools directly into software users already have installed.

That said, simultaneous interpretation is notoriously unforgiving. For structurally divergent language pairs like English and German or Japanese, translating just seconds behind the speaker leaves very little margin for error before context shifts entirely.