Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking this week—two real-time voice models that push voice AI beyond simple conversation.
The Extended Thinking mode is the standout. You give it a complex task, it says “let me check on that…”, and then it works in the background while reporting progress in real time. It’s not waiting until it’s finished to tell you what happened. It’s thinking out loud while it works, which makes the interaction feel far more natural than the “wait in silence, then get a result” pattern of previous voice models.
The benchmarks back up the hype:
- Speech to Speech Quality Index: 82.6, ranked #1 overall
- τ-Voice agentic task completion: 68.6%, ranked #1
- Big Bench Audio reasoning: 97.7%
It supports 97 languages with automatic detection and switching, so you can change languages mid-conversation and it keeps up. All generated audio includes SynthID invisible watermarking, so AI-generated speech can be detected and traced.
My take: Voice AI has been stuck in the “demo-able but not useful” phase for a while. The Extended Thinking mode is the first feature that makes voice feel like a genuine workflow tool rather than a novelty. If you use voice with AI regularly, this is worth trying.
Laat een reactie achter