Google Gemini 3.8 Live Voice Model

Overview

Gemini 3.8 Live is Google’s real-time, audio-first dialogue model for natural voice conversations. It listens to live 16 kHz PCM microphone audio and streams back 24 kHz PCM speech while the session stays open. It can also accept text, images, and video as context, then answer in audio and text. This model helps you build assistants that feel like a phone call, not a chat box.

What you can build

• A customer support voice agent that answers, clarifies, and confirms details out loud
• A hands-free assistant for field technicians who talk while working
• A tutoring companion that explains steps in spoken form and handles follow-up questions
• A voice concierge that can browse current facts with Search grounding when enabled
• A multimodal voice helper that reacts to a camera feed or screenshots while speaking

Inputs

• A required voice selection for the assistant’s speech, chosen from a fixed set of prebuilt voices. The model automatically speaks the conversation language.
• Required system instructions (plain text) that define tone, style, and boundaries for the assistant.
• An optional thinking depth setting. When you choose a non-none level, the run switches to the Extended Thinking variant for more background reasoning during live dialogue.
• An optional toggle to allow Google Search grounding for current information.
• Live microphone audio as raw 16-bit PCM (little-endian). A 16 kHz sample rate is the native and recommended format, with optional 8 kHz or 24 kHz input.
• Optional speech start sensitivity that controls how easily the system treats sound as the start of speech.
• Optional speech end sensitivity that controls how quickly the system ends your turn after you pause.
• Optional silence duration (milliseconds) that sets how long you must be quiet before the assistant starts speaking.

Outputs

• Streamed spoken responses as raw 16-bit PCM audio (little-endian) at 24 kHz.
• Optional text transcripts of what you said, if input transcription is enabled by the platform.
• Optional text alongside speech, such as short summaries or confirmations, depending on the experience design.

Recommended settings

• Use 16 kHz, 16-bit PCM microphone input for best compatibility with the model’s native audio path.
• Keep thinking set to none for the lowest-latency back-and-forth.
• Switch thinking to low, medium, or high when you need multi-step reasoning during a live call.
• Use high speech end sensitivity when you want faster turn-taking.
• Lower speech start sensitivity in noisy rooms to reduce false triggers.

Limitations

• Knowledge cutoff: January 2025. Use Search grounding if you need current facts.
• Input limit is about 128K tokens and output limit is about 64K tokens, across supported modalities.
• Structured outputs aren’t supported, so you can’t rely on strict JSON-only responses.
• Context caching isn’t supported.
• Code execution and file search aren’t supported.
• Audio quality matters. Background noise, overlapping speech, and clipped microphones can reduce accuracy and timing.
• Non-structured visual inputs can fail. Blurry photos, tiny text, or shaky video can cause wrong conclusions.

Safety & compliance

• Follow Google’s Generative AI policies and the Gemini API terms for prohibited content and misuse.
• Don’t use the model to generate or facilitate harmful, illegal, or privacy-invasive content.
• Treat live audio as sensitive data. Get consent before recording, storing, or sharing user speech.
• Validate important claims. Like other foundation models, it can hallucinate or misunderstand context.