Google has introduced improvements to the Gemini 2.5 Flash Native Audio model, which focuses on better voice conversation processing. The model can now handle complex workflows, better understand user instructions, and conduct more natural dialogues. In applications such as Google AI Studio or Vertex AI, the model can call external functions more reliably to obtain up-to-date information during a conversation and incorporate it into its response without interruption. It achieved a score of 71.5% in the ComplexFuncBench Audio test, the best result among comparable models.
The model now follows instructions more reliably, with a success rate of 90% compared to the previous 84%, and has also improved at maintaining context across multiple rounds of conversation. This means that when you talk to Gemini Live or Search Live, the responses are smoother and more closely related to what you said previously. For example, in Search Live you can now brainstorm ideas together in real time, as if you were talking to an expert, and the model remembers details from previous questions.
Live speech translation directly through headphones
One of the biggest new features is live speech translation, which Gemini is bringing to the Google Translate app. This feature allows real-time translations to be streamed directly to headphones while preserving the speaker's intonation, pace, and pitch. It works in two modes: continuous listening, where the model translates speech from multiple languages into a single target language so that you hear the world around you in your own language, or two-way conversation, where the translation switches automatically depending on who is speaking.
For example, if you speak English and want to chat with someone who speaks Hindi, you will hear the English translation through your headphones, while your phone plays the Hindi translation when you finish speaking. The model supports more than 70 languages, including Czech, and 2,000 language pairs, automatically detects the language, filters out background noise, and handles mixed languages in a single session without requiring you to change any settings.
This beta version is currently available in the Google Translate app on Android in the US, Mexico, and India, with support for iOS and additional countries coming in 2026. Simply connect your headphones, open the app, and tap "Live translate".
