Skip to main content

Gemini 3.8 Live: Voice AI That Thinks While It Talks

Google's Latest Voice Models Can Multitask Mid-Sentence

Imagine chatting with a voice assistant that doesn't make you wait while it searches the web or runs a task. That's the pitch behind Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two new models Google just rolled out. They're designed to keep the conversation flowing—no awkward pauses, no "hold on, I'm thinking."

Image

What Makes Them Different?

The headline feature: background tool and API calls. While you're talking, the AI can quietly fire off a web search, check a database, or trigger an action. You don't have to stop mid-sentence. The models also auto-detect 97 languages and can switch between them on the fly. Need visual positioning? They handle that too, and they'll keep you posted with little verbal cues like "Let me check..." so you know something's happening.

How Do They Stack Up?

Benchmarks are early, but promising. Gemini 3.8 Live Extended Thinking scored 82.6 on the Artificial Analysis Speech to Speech Quality Index. In the Speech Agent Arena, the standard Live model took second place. The Extended Thinking version notched 68.6% on the T-Voice test, 35.1% on the Sierra T-Voice benchmark, and a hefty 97.7% on Big Bench Audio.

Built for Developers, Tuned for Real Life

Google isn't keeping this to itself. The company has teamed up with platforms like Agora, LiveKit, Vercel, Pipecat, Fishjam, and Vision Agents to make integration smoother. You'll find the models rolling out on the Gemini API and Google AI Studio, with applications in Search Live, Gemini Enterprise, Google Workspace, and Gemini Live.

Worried about AI-generated audio being used for mischief? Google's got that covered. Every clip produced by these models carries an invisible SynthID watermark, so you can tell what's real and what's synthetic.

The Bigger Picture

This upgrade pushes voice AI beyond simple listening and speaking. It's moving toward continuous, multitasking agents that understand, reason, and get things done—all without breaking the conversational flow. In other words, the days of waiting for your digital assistant to catch up might finally be numbered.

Key Points

  • Gemini 3.8 Live and Live Extended Thinking enable near real-time reasoning, synchronized voice and thinking, and background tool/API calls.
  • Models support 97 languages with automatic detection and mid-conversation switching.
  • Benchmark highlights: 82.6 on Speech to Speech Quality Index, 97.7% on Big Bench Audio, second place in Speech Agent Arena.
  • Developer integrations with Agora, LiveKit, Vercel, Pipecat, Fishjam, and Vision Agents.
  • All generated audio includes an invisible SynthID watermark for authenticity.