Google has published detailed technical documentation for the Gemini Live API, its interface for building real-time voice and vision agents on top of the Gemini model family. The reference page, hosted at ai.google.dev, specifies a stateful WebSocket protocol that carries continuous audio, image, and text streams, a different architecture than the request-response pattern used by Gemini’s standard text endpoint. That choice explains the target use cases Google lists: shopping assistants, in-game non-player characters, and voice interfaces for robotics, smart glasses, and vehicles.
The technical specifications are unusually concrete for a docs page. Input arrives as raw 16-bit PCM audio at 16kHz, JPEG images capped at one frame per second, and plain text. Output returns as raw 16-bit PCM audio at 24kHz. Google sets the spoken-language ceiling at 70 languages, and offers a separate Live Translation feature for real-time voice-to-voice translation across more than 70 languages.
Builders choose between two integration paths, and the choice has consequences beyond a single API call. A server-to-server model routes a user’s audio and video through a developer’s own backend before it reaches the Live API, preserving a point for logging, moderation, and auth. A client-to-server model connects a frontend directly over WebSockets using ephemeral tokens, trimming a network hop but pushing more trust onto code running in the browser. Google also documents a third route through its Agent Development Kit, which wraps the same streaming connection for teams that want an agent framework rather than raw WebSocket handling, plus partner integrations that support the Live API over WebRTC. Teams should settle on a path early: switching later means rearchitecting the data path, not swapping a single endpoint.
Google also lists barge-in, which lets a user interrupt the model mid-response, function calling and Google Search as callable tools, live transcription of both sides of the conversation, a “proactive audio” setting that gives developers control over when and in what contexts the model responds, and an “affective dialog” mode that adjusts tone to match how the user sounds. What the documentation does not include is any latency benchmark, Live-API-specific pricing, or a stated session-length limit, gaps that matter to anyone sizing a production voice product before committing engineering time.
Real-time voice has become a contested layer among the frontier labs, and a WebSocket-native design commits developers earlier to a streaming architecture than a standard chat completion call would. Teams evaluating the Live API should prototype both integration paths before choosing one and confirm current pricing and rate limits directly with Google, since neither figure appears on this reference page.
Google published these specifications in its Gemini API documentation for the Live API, last updated June 12, 2026.
