Google DeepMind developer-experience engineer Thor Schaeff says native-audio models could make voice agents more natural and responsive by handling speech directly, rather than routing it through text. His conference demonstration also showed how stateful APIs can carry context between generations, a useful step towards interactive AI that can keep a story, image or task coherent.
Google DeepMind Watch analysis
What happened
At the AI Engineer conference, Schaeff demonstrated Gemini 3.1 Flash Live, which he described as a native-audio model, and showed it identifying an object, reading text, translating and conversing in another language. The BigGo Finance account of his talk says the preview supports 90 languages, interruption mid-response and tool calls. It also reports that developers can access it through Google’s Live API, a stateful WebSocket interface for streaming text, audio and video.
Schaeff also presented Google’s Interactions API as a way to preserve context across model calls. In a story-and-image demo, successive generations carried forward the earlier interaction state, helping maintain consistency between images. The report says image generation took roughly two to three seconds per frame in that demonstration.
The talk covered three distinct audio offerings: Gemini TTS for expressive speech generation, Gemini 3.5 Live Translate for near-real-time translation, and Gemini 3.1 Flash Live for conversation. These are different tools for different jobs, not one voice model that magically does everything.
Why it matters
Removing a speech-to-text-to-speech chain could make conversations feel quicker and less like a relay race between models. Stateful calls could also spare developers from repeatedly restating characters, styles or task context. Those are meaningful building blocks for live applications, from translation to interactive media.
There is a practical gap between a convincing stage demonstration and a dependable service at scale. Schaeff pointed developers towards infrastructure partners LiveKit and Pipecat for production deployment, acknowledging that real-time media is still difficult to scale.
Our read
The useful news is the direction: native audio for live conversation, and persistent state for linked generations. The demos make that direction tangible, but they are not proof that every component is ready for demanding production use. Developers should treat the APIs as promising building blocks and test latency and reliability in their own workloads before putting them in front of customers. A demo can skip the queue; production cannot.
What to watch
- Whether the preview’s language support and interruption features hold up outside demonstrations.
- How developers use the Live API and Interactions API in shipped applications.
- Whether the infrastructure needed to run real-time agents reliably becomes simpler.
Discussion spark: For a voice agent, would you prioritise more natural, interruption-friendly conversation, or the reliability of a simpler text-based pipeline?
Sources and evidence
- Google DeepMind's Thor Schaeff: Native Audio Models Will Replace Cascading Voice Agents – BigGo Finance (11 October 2026, 17:08 UTC)
not affiliated with, endorsed by, or operated by Google or Google DeepMind