Google has announced a major upgrade to its audio-intelligence ecosystem, introducing the Gemini 2.5 Flash Native Audio model. Designed to power the next generation of live voice agents and real-time translation, these updates focus on making AI interactions feel less like a command-line interface and more like a natural human conversation.
The updated model is now rolling out across Google AI Studio, Vertex AI, Gemini Live, and—for the first time—Search Live.
A New Standard for Live Voice Agents
While text-to-speech has historically focused on “sounding” human, the new Gemini 2.5 Flash Native Audio focuses on “thinking” and “acting” human during live interactions. Google has optimized the model in three critical areas:
- Precision Function Calling: The model is significantly more reliable at “reaching out” to the internet or external apps. It can fetch real-time data mid-sentence and weave it into an audio response without awkward pauses. It currently leads the ComplexFuncBench Audio evaluation with a 71.5% score.
- Superior Instruction Following: For developers, the model now boasts a 90% adherence rate to complex instructions, ensuring voice agents stay on brand and on task.
- Contextual Memory: Multi-turn conversations are now smoother. The model is better at remembering what was said minutes ago, allowing for cohesive, long-form brainstorming or troubleshooting.
“Users often forget they’re talking to AI within a minute of using Sidekick… New Live API capabilities empower our merchants to win.”
—David Wurtz, VP of Product, Shopify
Breaking Language Barriers: Live Speech Translation
Perhaps the most ambitious update is the introduction of live speech-to-speech translation, currently rolling out in beta for the Google Translate app.
Unlike traditional translators that feel robotic, Gemini’s native audio capabilities enable Style Transfer. This means the translation preserves the original speaker’s intonation, pacing, and pitch. If the speaker sounds excited, the translation will too.
Key Translation Features:
- Continuous Listening: Wear headphones and hear the world around you translated into your primary language in real-time.
- Bilingual Flow: The model automatically detects which of two languages is being spoken and switches the output instantly.
- Noise Robustness: Designed for the real world, the model can filter out ambient street noise or cafe chatter to focus on the speaker.
Availability and Rollout
The Gemini 2.5 Flash Native Audio model is now generally available on Vertex AI and in preview for the Gemini API.
The Live Translate beta experience is rolling out today on Android devices in the US, Mexico, and India, with iOS support and further regions expected shortly. Google plans to bring these advanced translation capabilities to the broader Gemini API for developers in 2026.