Google has made available two new speech-to-speech models today: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. The models follow a similar approach to OpenAI’s GPT-Live family, enabling real-time reception of spoken input and generation of spoken responses.
Browser-based test UI built by a developer
A developer, who used GPT-6 Astra Extra High to parse the documentation, built a browser user interface for trying the new models. The UI lets users:
- select a model and a voice preset,
- provide an optional system prompt,
- start a voice conversation directly in the browser,
- interrupt the model while it is speaking.
Implementation details
According to the developer, the implementation uses no external libraries. It connects to Google’s generative language WebSocket endpoint at wss://generativelanguage.googleapis.com/ws/google.ai.generativelanguage.v1alpha.GenerativeService.BidiGenerateContent?key=... and employs the Web Audio API’s AudioContext for both capture and playback.
Why this matters
The Gemini 3.8 Live models represent another step toward real-time speech processing: applications can more easily add in-browser, interruptible voice interactions. The WebSocket-based, library-free approach simplifies prototype development for engineers who want to quickly test the models’ capabilities.
Limitations and notes
The demonstration UI and its settings are an individual developer’s implementation; not all details are publicly documented. The included WebSocket URL points to Google’s generative language API and the ?key=... parameter indicates that an API key is required to use the endpoint.
