Gemini STT

Google’s Gemini provides streaming speech-to-text through its Live transcription models, with automatic language detection and built-in turn detection.

Vision Agents uses Stream Video for real-time WebRTC transport by default. External WebRTC transports are supported as well. Most AI providers offer free tiers to get started.

Gemini also provides an LLM with built-in tools and Realtime speech-to-speech. You can use the STT and LLM in the same agent.

Installation

uv add "vision-agents[gemini]"

Quick Start

from vision_agents.core import Agent, User
from vision_agents.plugins import elevenlabs, gemini, getstream

agent = Agent(
    edge=getstream.Edge(),
    agent_user=User(name="Assistant", id="agent"),
    instructions="You are a helpful assistant.",
    llm=gemini.LLM(),
    stt=gemini.STT(),
    tts=elevenlabs.TTS(),
)

Set GOOGLE_API_KEY or GEMINI_API_KEY in your environment, or pass api_key directly.

Parameters

stt = gemini.STT(
    language_codes=["en-US"],
    custom_vocabulary=["Vision Agents", "GetStream"],
)

Name	Type	Default	Description
`model`	`str`	`"gemini-3.5-transcribe-live"`	Gemini Live transcription model
`language_codes`	`list[str]`	`None`	BCP-47 language codes. Omit for automatic language detection
`custom_vocabulary`	`list[str]`	`None`	Phrases that should be recognized accurately
`http_options`	`HttpOptions`	`None`	Gemini HTTP configuration
`client`	`genai.Client`	`None`	Preconfigured Gemini client
`api_key`	`str`	`None`	API key (defaults to `GOOGLE_API_KEY` or `GEMINI_API_KEY` env var)

Language Detection

Gemini detects the spoken language automatically. Pass language_codes when you know which languages to expect, which helps in multilingual calls:

stt = gemini.STT(language_codes=["en-US", "de-DE"])

Custom Vocabulary

Product names, jargon, and proper nouns are often transcribed incorrectly. Pass them as custom_vocabulary to bias recognition towards the spelling you want:

stt = gemini.STT(
    custom_vocabulary=["Vision Agents", "GetStream", "Mintlify"],
)

Transcripts and Turns

The plugin streams 16 kHz PCM audio to Gemini and emits standard transcript events as speech arrives, followed by a final transcript when the utterance completes. Turn detection is built in, so no separate turn detection plugin is required.

Google also publishes gemini-3.5-transcribe for file transcription. This plugin uses the Live model for real-time streaming.

Overview

Language Models

Realtime

Speech-to-Text

Text-to-Speech

Vision & Video

Avatars

Turn Detection

Infrastructure

Edge Transport

Telephony

Custom Integrations

Installation

Quick Start

Parameters

Language Detection

Custom Vocabulary

Transcripts and Turns

Next Steps

Gemini LLM

Build a Voice Agent

​Installation

​Quick Start

​Parameters

​Language Detection

​Custom Vocabulary

​Transcripts and Turns

​Next Steps

Gemini LLM

Build a Voice Agent

Installation

Quick Start

Parameters

Language Detection

Custom Vocabulary

Transcripts and Turns

Next Steps