Google introduces Gemini 3.5 Live Translate for near-real-time, natural speech translation

Google is rolling out a new capability for live speech translation with Gemini 3.5 Live Translate, positioning it as a shift toward more fluid, near-real-time conversations across languages. Announced in a Google engineering blog post, the company describes the system as an audio model built for speech-to-speech translation, rather than a turn-by-turn workflow that waits for a speaker to finish before responding.

In Google’s description, Gemini 3.5 Live Translate can automatically detect 70-plus languages and produce speech that it says sounds natural, including the speaker’s intonation, pacing and pitch. The company contrasts the approach with “turn by turn” translation systems, arguing that waiting for the entire utterance can make real-time interaction feel choppy.

A core design goal, according to the post, is to generate translated speech continuously as audio is streamed into the system. Google says this creates smoother output without awkward pauses, while the translated audio stays only a few seconds behind the speaker throughout the session. It also says the model processes speech as it is streamed, supporting a more seamless connection across languages.

Google says the model is built to handle real-world conditions, including noisy environments. It describes noise robustness as a feature for deployments such as multilingual calls, meetings, lessons and broadcasts, where audio can vary and background sound is unpredictable. The company also says multilingual inputs can be handled without requiring users to manually configure settings.

The company plans to make Gemini 3.5 Live Translate available in several Google products, including Google AI Studio, Google Translate and Google Meet. For consumers, that could mean closer-to-live translation in environments where conversation timing matters, while for developers it expands tooling for building interpretation-style experiences.

For developers, Google points to the Gemini Live API, which it describes as enabling dubbing and simultaneous multi-language translation. The post frames this as a way to integrate continuous, live translation into third-party applications, including collaboration tools and services that support multilingual participants.

Google also links to demonstrations and example code in the Gemini Cookbook, using the API as the development pathway for dubbing and simultaneous translation workflows. In addition, the company highlights use cases ranging from live interpretation for calls and meetings to supporting multilingual broadcast and other real-time scenarios.

Still, the announcement does not provide several details that typically matter for enterprise rollouts. Google does not specify latency beyond the general description that translated audio stays a few seconds behind, does not disclose cost, availability dates, or regional rollout plans, and does not outline how quality is evaluated across language pairs. The post also does not describe what data handling or privacy controls apply to speech inputs in these products beyond pointing developers to the API and sample materials.