Google’s new speech model Gemini 3.8 Live supports real-time reasoning

Google has unveiled Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, its latest voice processing AI models, aimed at tackling latency issues in voice-based AI agents. These models are heralded as the most advanced AI audio systems the company has released thus far, boasting capabilities for real-time reasoning and simultaneous speech and thought processing.

They also have the ability to perform background tasks, such as executing third-party software tool calls, allowing for a more natural, uninterrupted user experience.

Gemini 3.8 Live Extended Thinking, in particular, has set a new benchmark with a score of 82.6 on the Artificial Analysis Speech to Speech Quality Index, outperforming previous models like GPT-Live-1-Astra and Grok Voice Think Fast 2.0. It achieved a score of 68.6% on T-Voice, 35.1% on T-Voice-banking, and a remarkable 97.7% on Big Bench Audio.

The key difference between the two models lies in their functionality. While Gemini 3.8 Live is available now for use via the Gemini API, Google AI Studio, and Google Workspace, as well as through Google partner platforms, Gemini 3.8 Live Extended Thinking offers an additional layer of reasoning capabilities, available for a fee based on the number of reasoning tokens used and other inputs.

Both models support automatic language detection and can switch between 97 languages mid-conversation, providing speech generation in these languages. They also support near real-time visual grounding and can acknowledge verbal prompts like "let me check that."