Google DeepMind shipped Gemini 3.8 Live and an Extended Thinking variant on September 15 — speech-to-speech models that detect a language switch mid-conversation and run tasks in the background without interrupting speech. Extended Thinking takes first place on the Speech to Speech Quality Index with a score of 82.6.
Key takeaways
- Extended Thinking: 1st on the Speech to Speech Quality Index (82.6 points)
- Big Bench Audio: 97.7% on reasoning over audio material
- Agentic tasks: 68.6% on τ-Voice and 35.1% on τ-Voice-banking
- Language detection and switching mid-conversation across 97 languages
- All generated audio carries SynthID watermarking
A model that thinks and talks at once
The core difference between the variants is what happens while the model is speaking. Gemini 3.8 Live Extended Thinking reasons in parallel with speaking rather than pausing to think an answer through. Both models make tool and API calls in the background without interrupting the dialogue.
They detect a language change mid-conversation and switch to the new one smoothly, across 97 supported languages. Visual input is processed in near real time. This is classic speech-to-speech AI: audio in, audio out, with no intermediate transcription. All generated audio is watermarked with SynthID.
What the benchmarks say
| Test | Result | What it measures |
|---|---|---|
| Speech to Speech Quality Index | 82.6 — 1st place | voice conversation quality |
| Big Bench Audio | 97.7% | reasoning over audio material |
| τ-Voice | 68.6% | agentic tasks completed by voice |
| τ-Voice-banking | 35.1% | banking tasks handled by voice |
The Speech to Speech Quality Index is run by Artificial Analysis. Gemini 3.8 Live itself took second place in the Speech Agent Arena. The weakest results come from voice-driven agentic tasks, especially the banking variant of τ-Voice from Sierra.
First impressions from the API
Simon Willison built a browser interface for testing both models on launch day — with no libraries at all. It connects over WebSocket and uses the Web Audio API for capture and playback. The interface lets you pick the model variant, the voice and a system prompt, and interrupt the model while it is talking.
// The bidirectional voice-session endpoint the interface talks to
const ws = new WebSocket(
'wss://generativelanguage.googleapis.com/ws/' +
'google.ai.generativelanguage.v1alpha.GenerativeService.BidiGenerateContent'
);
// Microphone capture and response playback are handled by the Web Audio API —
// with no external library at all.The code went into his public tools repository.
Why it matters
Mid-conversation language detection removes a step that until now needed a separate classifier or an explicit user choice. Running tools in the background eliminates the silence that gives away a machine on the other end. Those are the two things that separate a demo from a product. The 35.1% on banking τ-Voice shows where the limit sits, though.
What's next?
- For developers the models are available in the Gemini API and Google AI Studio
- Google DeepMind disclosed no pricing — the announcement refers only to a competitive cost level
- Gemini Enterprise for Customer Experience has been announced without a date, and the enterprise version sits in private preview
Sources
- Google — Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking
- Simon Willison — Gemini Live Audio
