VAD Metrics | LiveKit Documentation

This example shows you how to log voice-activity-detection (VAD) metrics during a call. Each time the Silero VAD processes speech, it emits idle time and inference timing data that you render with Rich.

Note

This recipe uses the per-plugin metrics_collected event on the VAD instance. This per-component surface is not deprecated. A separate session-level metrics_collected event (session.on("metrics_collected", ...)) is deprecated. For session-scoped cost and usage tracking, see Session usage.

Prerequisites

Add a .env in this directory with your LiveKit credentials:

LIVEKIT_URL=your_livekit_url
LIVEKIT_API_KEY=your_api_key
LIVEKIT_API_SECRET=your_api_secret

Install dependencies:

pip install rich "livekit-agents[silero]" python-dotenv

Load environment, logging, and define an AgentServer

Set up dotenv, logging, a Rich console for the VAD reports, and initialize the AgentServer.

import logging
import asyncio
from dotenv import load_dotenv
from livekit.agents import JobContext, JobProcess, AgentServer, cli, Agent, AgentSession, inference
from livekit.agents.metrics import VADMetrics
from livekit.plugins import silero
from rich.console import Console
from rich.table import Table
from rich import box
from datetime import datetime

load_dotenv()

logger = logging.getLogger("metrics-vad")
logger.setLevel(logging.INFO)

console = Console()

server = AgentServer()

Define a lightweight agent and VAD metrics display function

Keep the Agent class minimal with just instructions. Define an async function to display VAD metrics as a Rich table.

class VADMetricsAgent(Agent):
    def __init__(self) -> None:
        super().__init__(
            instructions="You are a helpful agent."
        )

async def display_vad_metrics(metrics: VADMetrics):
    table = Table(
        title="[bold blue]VAD Metrics Report[/bold blue]",
        box=box.ROUNDED,
        highlight=True,
        show_header=True,
        header_style="bold cyan"
    )

    table.add_column("Metric", style="bold green")
    table.add_column("Value", style="yellow")

    timestamp = datetime.fromtimestamp(metrics.timestamp).strftime('%Y-%m-%d %H:%M:%S')

    table.add_row("Type", str(metrics.type))
    table.add_row("Label", str(metrics.label))
    table.add_row("Timestamp", timestamp)
    table.add_row("Idle Time", f"[white]{metrics.idle_time:.4f}[/white]s")
    table.add_row("Inference Duration Total", f"[white]{metrics.inference_duration_total:.4f}[/white]s")
    table.add_row("Inference Count", str(metrics.inference_count))

    console.print("\n")
    console.print(table)
    console.print("\n")

Prewarm VAD for faster connections

Preload the VAD model once per process. This runs before any sessions start and stores the VAD instance in proc.userdata so it can be reused, cutting down on connection latency.

def prewarm(proc: JobProcess):
    proc.userdata["vad"] = silero.VAD.load()

server.setup_fnc = prewarm

Define the rtc session with VAD metrics hook

Create an rtc session entrypoint that retrieves the prewarmed VAD, hooks into its metrics_collected event, and starts the agent session with STT/LLM/TTS configuration.

@server.rtc_session(agent_name="my-agent")
async def entrypoint(ctx: JobContext):
    ctx.log_context_fields = {"room": ctx.room.name}

    vad_instance = ctx.proc.userdata["vad"]

    def on_vad_metrics(metrics: VADMetrics):
        asyncio.create_task(display_vad_metrics(metrics))

    vad_instance.on("metrics_collected", on_vad_metrics)

    session = AgentSession(
        stt=inference.STT(model="deepgram/nova-3-general"),
        llm=inference.LLM(model="openai/gpt-5-mini"),
        tts=inference.TTS(model="cartesia/sonic-3", voice="9626c31c-bec5-4cca-baa8-f8ba9e84c8bc"),
        vad=vad_instance,
        preemptive_generation=True,
    )

    await session.start(agent=VADMetricsAgent(), room=ctx.room)
    await ctx.connect()

Run the server

The cli.run_app() function starts the agent server. It manages the worker lifecycle, connects to LiveKit, and processes incoming jobs.

if __name__ == "__main__":
    cli.run_app(server)

Run it

python metrics_vad.py console

How it works

The VAD model is prewarmed once per process for faster connections.
When the rtc session starts, the metrics_collected event handler is attached to the VAD.
Silero VAD detects speech and emits metrics events with idle time, inference duration, and count.
A background task formats and prints the metrics as a Rich table.
Because the handler is async, it does not block ongoing audio processing.

Full example

import logging
import asyncio
from dotenv import load_dotenv
from livekit.agents import JobContext, JobProcess, AgentServer, cli, Agent, AgentSession, inference
from livekit.agents.metrics import VADMetrics
from livekit.plugins import silero
from rich.console import Console
from rich.table import Table
from rich import box
from datetime import datetime

load_dotenv()

logger = logging.getLogger("metrics-vad")
logger.setLevel(logging.INFO)

console = Console()

class VADMetricsAgent(Agent):
    def __init__(self) -> None:
        super().__init__(
            instructions="You are a helpful agent."
        )

async def display_vad_metrics(metrics: VADMetrics):
    table = Table(
        title="[bold blue]VAD Metrics Report[/bold blue]",
        box=box.ROUNDED,
        highlight=True,
        show_header=True,
        header_style="bold cyan"
    )

    table.add_column("Metric", style="bold green")
    table.add_column("Value", style="yellow")

    timestamp = datetime.fromtimestamp(metrics.timestamp).strftime('%Y-%m-%d %H:%M:%S')

    table.add_row("Type", str(metrics.type))
    table.add_row("Label", str(metrics.label))
    table.add_row("Timestamp", timestamp)
    table.add_row("Idle Time", f"[white]{metrics.idle_time:.4f}[/white]s")
    table.add_row("Inference Duration Total", f"[white]{metrics.inference_duration_total:.4f}[/white]s")
    table.add_row("Inference Count", str(metrics.inference_count))

    console.print("\n")
    console.print(table)
    console.print("\n")

server = AgentServer()

def prewarm(proc: JobProcess):
    proc.userdata["vad"] = silero.VAD.load()

server.setup_fnc = prewarm

@server.rtc_session(agent_name="my-agent")
async def entrypoint(ctx: JobContext):
    ctx.log_context_fields = {"room": ctx.room.name}

    vad_instance = ctx.proc.userdata["vad"]

    def on_vad_metrics(metrics: VADMetrics):
        asyncio.create_task(display_vad_metrics(metrics))

    vad_instance.on("metrics_collected", on_vad_metrics)

    session = AgentSession(
        stt=inference.STT(model="deepgram/nova-3-general"),
        llm=inference.LLM(model="openai/gpt-5-mini"),
        tts=inference.TTS(model="cartesia/sonic-3", voice="9626c31c-bec5-4cca-baa8-f8ba9e84c8bc"),
        vad=vad_instance,
        preemptive_generation=True,
    )

    await session.start(agent=VADMetricsAgent(), room=ctx.room)
    await ctx.connect()

if __name__ == "__main__":
    cli.run_app(server)