Google推出Gemini 3.8 Live系列模型，性能领先语音对话领域

Google推出了Gemini 3.8 Live和Gemini 3.8 Live Extended Thinking两款新型AI模型，专为语音助手和实时对话场景设计1。其中，Gemini 3.8 Live Extended Thinking在Artificial Analysis语音质量指数中获得82.6分，排名第一，并在agentic任务完成度上表现突出，τ-Voice基准达到68.6%、Sierra's τ-Voice-banking基准达35.1%1。Gemini 3.8 Live在Speech Agent Arena中排名第二1。

两款模型具备广泛的多语言支持与安全特性1。它们能够自动检测并实时切换97种语言，所有AI生成的音频均采用SynthID水印技术进行标记1。Big Bench Audio基准测试中该系列模型取得97.7%的得分1。

两款模型已开始向不同用户群体推出1。Gemini 3.8 Live向开发者（通过Gemini API和Google AI Studio）、企业（Gemini Enterprise私密预览）和普通用户（Search Live）推出1；Gemini 3.8 Live Extended Thinking向开发者、企业以及Google AI Pro和Ultra订阅用户推出1。

Google has unveiled two new AI models optimized for voice assistants and real-time conversation: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking 1. The Gemini 3.8 Live Extended Thinking model achieved the top ranking on Artificial Analysis' Speech to Speech Quality Index with a score of 82.6 1. On the τ-Voice benchmark, it demonstrated agentic task completion performance of 68.6%, with 35.1% on Sierra's τ-Voice-banking benchmark 1. Gemini 3.8 Live ranked second in the Speech Agent Arena, with both models scoring 97.7% on the Big Bench Audio benchmark 1.

Both models are now rolling out to developers through the Gemini API and Google AI Studio, to enterprises via Gemini Enterprise private preview, and to general users through Search Live 1. Gemini 3.8 Live Extended Thinking is additionally available to Google AI Pro and Ultra subscription users 1. The models support real-time automatic detection and switching across 97 languages, and all AI-generated audio is marked with SynthID watermarking technology 1.