Realtime Voice Agents

Realtime agents enable voice-based interactions with AI assistants using bidirectional audio streaming. The adk-realtime crate provides a unified interface for building voice-enabled agents that work with OpenAI's Realtime API and Google's Gemini Live API.

Overview

Realtime agents differ from text-based LlmAgents in several key ways:

Architecture

              ┌─────────────────────────────────────────┐
              │              Agent Trait                │
              │  (name, description, run, sub_agents)   │
              └────────────────┬────────────────────────┘
                               │
       ┌───────────────────────┼───────────────────────┐
       │                       │                       │
┌──────▼──────┐      ┌─────────▼─────────┐   ┌─────────▼─────────┐
│  LlmAgent   │      │  RealtimeAgent    │   │  SequentialAgent  │
│ (text-based)│      │  (voice-based)    │   │   (workflow)      │
└─────────────┘      └───────────────────┘   └───────────────────┘

RealtimeAgent implements the same Agent trait as LlmAgent, sharing:

• Instructions (static and dynamic)
• Tool registration and execution
• Callbacks (before_agent, after_agent, before_tool, after_tool)
• Sub-agent handoffs

Quick Start

Installation

Add to your Cargo.toml:

[dependencies]
adk-realtime = { version = "2.0.0", features = ["openai"] }

# For Vertex AI Live (Google Cloud with ADC auth)
# adk-realtime = { version = "2.0.0", features = ["vertex-live"] }

# For LiveKit WebRTC bridge
# adk-realtime = { version = "2.0.0", features = ["livekit"] }

# For all transports (except WebRTC which needs cmake)
# adk-realtime = { version = "2.0.0", features = ["full"] }

Basic Usage

use adk_realtime::{
    RealtimeAgent, RealtimeModel, RealtimeConfig, ServerEvent,
    openai::OpenAIRealtimeModel,
};
use std::sync::Arc;

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let api_key = std::env::var("OPENAI_API_KEY")?;

    // Create the realtime model
    let model: Arc<dyn RealtimeModel> = Arc::new(
        OpenAIRealtimeModel::new(&api_key, "gpt-realtime")
    );

    // Build the realtime agent
    let agent = RealtimeAgent::builder("voice_assistant")
        .model(model.clone())
        .instruction("You are a helpful voice assistant. Be concise.")
        .voice("alloy")
        .server_vad()  // Enable voice activity detection
        .build()?;

    // Or use the low-level session API directly
    let config = RealtimeConfig::default()
        .with_instruction("You are a helpful assistant.")
        .with_voice("alloy")
        .with_modalities(vec!["text".to_string(), "audio".to_string()]);

    let session = model.connect(config).await?;

    // Send text and get response
    session.send_text("Hello!").await?;
    session.create_response().await?;

    // Process events
    while let Some(event) = session.next_event().await {
        match event? {
            ServerEvent::TextDelta { delta, .. } => print!("{}", delta),
            ServerEvent::AudioDelta { delta, .. } => {
                // Play audio (delta is base64-encoded PCM)
            }
            ServerEvent::ResponseDone { .. } => break,
            _ => {}
        }
    }

    Ok(())
}

Supported Providers

Note: gpt-realtime is OpenAI's latest realtime model with improved speech quality, emotion, and function calling capabilities.

Transport Options

ADK-Realtime supports multiple transport layers:

• WebSocket (default): Direct connection to OpenAI or Gemini. Simple, low-latency, works everywhere.
• Vertex AI Live: Connects to Gemini via Google Cloud with OAuth2 authentication (Application Default Credentials). Use when you need enterprise auth and GCP integration.
• LiveKit WebRTC: Production-grade WebRTC bridge. Routes audio through a LiveKit server for scalable, multi-participant scenarios.
• OpenAI WebRTC: Direct WebRTC connection to OpenAI with Opus codec and data channels. Requires cmake for building the Opus C library.

RealtimeAgent Builder

The RealtimeAgentBuilder provides a fluent API for configuring agents:

let agent = RealtimeAgent::builder("assistant")
    // Required
    .model(model)

    // Instructions (same as LlmAgent)
    .instruction("You are helpful.")
    .instruction_provider(|ctx| format!("User: {}", ctx.user_name()))

    // Voice settings
    .voice("alloy")  // Options: alloy, coral, sage, shimmer, etc.

    // Voice Activity Detection
    .server_vad()  // Use defaults
    .vad(VadConfig {
        mode: VadMode::ServerVad,
        threshold: Some(0.5),
        prefix_padding_ms: Some(300),
        silence_duration_ms: Some(500),
        interrupt_response: Some(true),
        eagerness: None,
    })

    // Tools (same as LlmAgent)
    .tool(Arc::new(weather_tool))
    .tool(Arc::new(search_tool))

    // Sub-agents for handoffs
    .sub_agent(booking_agent)
    .sub_agent(support_agent)

    // Callbacks (same as LlmAgent)
    .before_agent_callback(|ctx| async { Ok(()) })
    .after_agent_callback(|ctx, event| async { Ok(()) })
    .before_tool_callback(|ctx, tool, args| async { Ok(None) })
    .after_tool_callback(|ctx, tool, result| async { Ok(result) })

    // Realtime-specific callbacks
    .on_audio(|audio_chunk| { /* play audio */ })
    .on_transcript(|text| { /* show transcript */ })

    .build()?;

Voice Activity Detection (VAD)

VAD enables natural conversation flow by detecting when the user starts and stops speaking.

Server VAD (Recommended)

let agent = RealtimeAgent::builder("assistant")
    .model(model)
    .server_vad()  // Uses sensible defaults
    .build()?;

Custom VAD Configuration

use adk_realtime::{VadConfig, VadMode};

let vad = VadConfig {
    mode: VadMode::ServerVad,
    threshold: Some(0.5),           // Speech detection sensitivity (0.0-1.0)
    prefix_padding_ms: Some(300),   // Audio to include before speech
    silence_duration_ms: Some(500), // Silence before ending turn
    interrupt_response: Some(true), // Allow interrupting assistant
    eagerness: None,                // For SemanticVad mode
};

let agent = RealtimeAgent::builder("assistant")
    .model(model)
    .vad(vad)
    .build()?;

Semantic VAD (Gemini)

For Gemini models, you can use semantic VAD which considers meaning:

let vad = VadConfig {
    mode: VadMode::SemanticVad,
    eagerness: Some("high".to_string()),  // low, medium, high
    ..Default::default()
};

Tool Calling

Realtime agents support tool calling during voice conversations:

use adk_realtime::{config::ToolDefinition, ToolResponse};
use serde_json::json;

// Define tools
let tools = vec![
    ToolDefinition {
        name: "get_weather".to_string(),
        description: Some("Get weather for a location".to_string()),
        parameters: Some(json!({
            "type": "object",
            "properties": {
                "location": { "type": "string" }
            },
            "required": ["location"]
        })),
    },
];

let config = RealtimeConfig::default()
    .with_tools(tools)
    .with_instruction("Use tools to help the user.");

let session = model.connect(config).await?;

// Handle tool calls in the event loop
while let Some(event) = session.next_event().await {
    match event? {
        ServerEvent::FunctionCallDone { call_id, name, arguments, .. } => {
            // Execute the tool
            let result = execute_tool(&name, &arguments);

            // Send the response
            let response = ToolResponse::new(&call_id, result);
            session.send_tool_response(response).await?;
        }
        _ => {}
    }
}

Multi-Agent Handoffs

Transfer conversations between specialized agents:

// Create sub-agents
let booking_agent = Arc::new(RealtimeAgent::builder("booking_agent")
    .model(model.clone())
    .instruction("Help with reservations.")
    .build()?);

let support_agent = Arc::new(RealtimeAgent::builder("support_agent")
    .model(model.clone())
    .instruction("Help with technical issues.")
    .build()?);

// Create main agent with sub-agents
let receptionist = RealtimeAgent::builder("receptionist")
    .model(model)
    .instruction(
        "Route customers: bookings → booking_agent, issues → support_agent. \
         Use transfer_to_agent tool to hand off."
    )
    .sub_agent(booking_agent)
    .sub_agent(support_agent)
    .build()?;

When the model calls transfer_to_agent, the RealtimeRunner handles the handoff automatically.

Audio Formats

use adk_realtime::{AudioFormat, AudioChunk};

// Create audio format
let format = AudioFormat::pcm16_24khz();

// Work with audio chunks
let chunk = AudioChunk::new(audio_bytes, format);
let base64 = chunk.to_base64();
let decoded = AudioChunk::from_base64(&base64, format)?;

Event Types

Server Events

Client Events

Vertex AI Live (Google Cloud)

Connect to Gemini Live via Vertex AI with enterprise authentication (ADC, service accounts, WIF):

use adk_realtime::gemini::{GeminiLiveBackend, GeminiRealtimeModel};
use adk_realtime::{RealtimeConfig, RealtimeModel};

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let project_id = std::env::var("GOOGLE_CLOUD_PROJECT")?;
    let region = std::env::var("GOOGLE_CLOUD_REGION")
        .unwrap_or_else(|_| "us-central1".to_string());

    // Use Application Default Credentials
    let credentials = google_cloud_auth::credentials::Builder::default()
        .build()
        .await?;

    let backend = GeminiLiveBackend::Vertex { credentials, region, project_id };
    let model = GeminiRealtimeModel::new(backend, "models/gemini-live-2.5-flash-native-audio");

    let config = RealtimeConfig::default()
        .with_instruction("You are a helpful voice assistant.");

    let session = model.connect(config).await?;
    session.send_text("Hello from Vertex AI!").await?;
    session.create_response().await?;

    // Process events...
    Ok(())
}

There's also a convenience constructor for ADC:

let model = GeminiRealtimeModel::vertex_adc(
    "us-central1",
    "my-project-id",
    "models/gemini-live-2.5-flash-native-audio",
).await?;

Vertex AI Live with Tool Calling

The vertex_live_tools example demonstrates function calling over a Vertex AI Live session:

use adk_realtime::config::ToolDefinition;
use adk_realtime::events::ToolResponse;
use serde_json::json;

// Declare tools
let tools = vec![
    ToolDefinition {
        name: "get_weather".to_string(),
        description: Some("Get current weather for a city".to_string()),
        parameters: Some(json!({
            "type": "object",
            "properties": {
                "city": { "type": "string" }
            },
            "required": ["city"]
        })),
    },
];

let config = RealtimeConfig::default()
    .with_tools(tools)
    .with_instruction("Use tools to answer questions about weather.");

let session = model.connect(config).await?;

// Handle FunctionCallDone events and send ToolResponse back
while let Some(event) = session.next_event().await {
    match event? {
        ServerEvent::FunctionCallDone { call_id, name, arguments, .. } => {
            let result = match name.as_str() {
                "get_weather" => json!({"temperature": "22°C", "condition": "sunny"}),
                _ => json!({"error": "unknown tool"}),
            };
            session.send_tool_response(ToolResponse::new(&call_id, result)).await?;
        }
        ServerEvent::TextDelta { delta, .. } => print!("{delta}"),
        ServerEvent::ResponseDone { .. } => break,
        _ => {}
    }
}

Feature Flags

LiveKit WebRTC Bridge

For production voice applications, the LiveKit bridge routes audio through a LiveKit server for scalable, multi-participant scenarios.

LiveKitConfig

Securely configure LiveKit credentials. API keys and secrets are stored using secrecy::SecretString and redacted in Debug output:

use adk_realtime::livekit::{LiveKitConfig, LiveKitRoomBuilder};

let config = LiveKitConfig::new(
    "wss://your-server.livekit.cloud",
    std::env::var("LIVEKIT_API_KEY")?,
    std::env::var("LIVEKIT_API_SECRET")?,
)?;

LiveKitConfig::new() validates the URL format and rejects empty credentials at construction time.

LiveKitRoomBuilder

A typestate builder for connecting to LiveKit rooms. The identity field is required at compile time — connect() is only available after it's set:

let bundle = LiveKitRoomBuilder::new(config)
    .identity("my-agent")           // required — enables connect()
    .name("Voice Agent")            // optional display name
    .room_name("session-room-123")  // optional — auto-generated if omitted
    .auto_subscribe(true)           // subscribe to remote tracks
    .with_audio(24_000, 1)          // publish a local audio track (sample rate, channels)
    .connect()
    .await?;

// The bundle contains everything you need
let room = bundle.room;
let mut events = bundle.events;
let audio_source = bundle.audio_source;  // for publishing audio
let audio_track = bundle.audio_track;

Bridging Audio

Use the bridge utilities to connect LiveKit audio to a RealtimeRunner:

use adk_realtime::livekit::{LiveKitEventHandler, bridge_input};

// Wrap your event handler to publish model audio to LiveKit
let lk_handler = LiveKitEventHandler::new(inner_handler, audio_source, 24000, 1);

// Bridge participant audio from LiveKit into the RealtimeRunner
tokio::spawn(bridge_input(remote_track, runner));

Examples

Run the included examples:

# OpenAI Realtime (WebSocket)
cargo run -p adk-realtime --example openai_session_update --features openai

# Vertex AI Live (requires gcloud auth application-default login)
cargo run -p adk-realtime --example vertex_live_voice --features vertex-live
cargo run -p adk-realtime --example vertex_live_tools --features vertex-live

# LiveKit Bridge (requires LiveKit server)
cargo run -p adk-realtime --example livekit_bridge --features livekit,openai
cargo run -p adk-realtime --example livekit_gemini_bridge --features livekit,gemini

# Debug utilities
cargo run -p adk-realtime --example debug_gemini --features gemini
cargo run -p adk-realtime --example debug_livekit_auth --features livekit

# OpenAI WebRTC (requires cmake)
cargo run -p adk-realtime --example openai_webrtc --features openai-webrtc

Best Practices

1. Use Server VAD: Let the server handle speech detection for lower latency
2. Handle interruptions: Enable interrupt_response for natural conversations
3. Keep instructions concise: Voice responses should be brief
4. Test with text first: Debug your agent logic with text before adding audio
5. Handle errors gracefully: Network issues are common with WebSocket connections

Comparison with OpenAI Agents SDK

ADK-Rust's realtime implementation follows the OpenAI Agents SDK pattern:

---

Previous: ← Graph Agents | Next: Model Providers →