What is the difference between vad and turn detection, which is better

Thanks for reaching out, Aditya. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.

how can i use them in my python livekit agent ?

You're asking how to use VAD and turn detection in a Python LiveKit agent. Below is a concise, runnable guide with examples for the common modes (VAD-only, STT endpointing, context-aware turn detector, and manual control). 1) Install

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

2) Prewarm VAD (recommended)

from dotenv import load_dotenv
from livekit.agents import AgentServer, JobProcess
from livekit.plugins import silero

load_dotenv()
server = AgentServer()

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

server.setup_fnc = prewarm

3) Example entrypoint patterns • VAD-only (silence endpointing)

from livekit.agents import AgentSession, AgentServer, cli, JobContext

@server.rtc_session()
async def entrypoint(ctx: JobContext):
    session = AgentSession(
        turn_detection="vad",
        vad=ctx.proc.userdata["vad"],
        stt="deepgram/nova-3-general",
        llm="openai/gpt-4.1-mini",
        tts="cartesia/sonic-3"
    )
    await session.start(agent=Agent(), room=ctx.room)
    await ctx.connect()

• STT endpointing (use provider phrase-end signals + VAD for interruptions)

from livekit.plugins import assemblyai

session = AgentSession(
    turn_detection="stt",
    stt=assemblyai.STT(),         # provider that supports phrase endpointing
    vad=ctx.proc.userdata["vad"]  # keeps interruption responsiveness
)

• Context-aware turn detection (recommended for natural conversations)

from livekit.plugins.turn_detector.multilingual import MultilingualModel

session = AgentSession(
    turn_detection=MultilingualModel(),  # or EnglishModel()
    vad=ctx.proc.userdata["vad"],
    stt="deepgram/nova-3-general",
    llm="openai/gpt-4.1-mini",
    tts="cartesia/sonic-3"
)

• Manual turn control (explicit programmatic control)

session = AgentSession(
    turn_detection="manual",
    vad=ctx.proc.userdata["vad"]
)

# later in code:
session.interrupt()
session.clear_user_turn()
session.commit_user_turn()

4) Tuning endpoint delays

session = AgentSession(
    turn_detection="vad",
    vad=ctx.proc.userdata["vad"],
    min_endpointing_delay=0.5,  # seconds after high-confidence end
    max_endpointing_delay=6.0   # seconds after low-confidence end
)

Notes and best practices • Always include

vad

even if using STT or realtime models—VAD gives low-latency interruption cues. • Use

MultilingualModel()

(or similar context-aware turn detectors) for natural conversational behavior; VAD-only is simpler but more prone to false cutoffs. • Prewarm the VAD model to reduce first-call latency. If you tell me your use case (IVR, voice assistant, meeting bot), language(s), and latency or compute limits, I’ll recommend which mode and exact parameter values to use. Sources: Turns overview | LiveKit docs | LiveKit turn detector plugin | LiveKit docs | examples/frontdesk | livekit/livekit_composite | DeepWiki | examples/voice_agents | livekit/livekit_composite | DeepWiki | OpenAI Realtime API plugin guide | LiveKit docs

Expected type 'Literal["stt", "vad", "realtime_llm", "manual"] | _TurnDetector | NotGiven', got 'MultilingualModel' instead

turn_detection=MultilingualModel(),

This line gives me type warning, why ?

Thanks — you're seeing a type warning because the

turn_detection

parameter's type annotation expects one of: • the string literals

"stt"

,

"vad"

,

"realtime_llm"

, or

"manual"

, • a

_TurnDetector

-typed object, or •

NotGiven

.

MultilingualModel()

is a turn detector implementation, but its runtime class may not be declared to satisfy the static type expected by the annotation, so your type checker (mypy / PyCharm / VSCode) warns. How to fix it • Import and use the proper TurnDetector interface/type if provided (preferred). For example, if there is a

_TurnDetector

or

TurnDetector

base type, annotate or construct against that. Example:

from livekit.plugins.turn_detector.multilingual import MultilingualModel
  from livekit.plugins.turn_detector import TurnDetector  # or the actual interface name

  td: TurnDetector = MultilingualModel()
  session = AgentSession(turn_detection=td, ...)

• If the library exposes a concrete accepted type name, use that type in your code so the checker recognizes it. • If an interface type is not exported or available, narrow the type checker with a cast:

from typing import cast
  session = AgentSession(turn_detection=cast(Any, MultilingualModel()), ...)

or

from typing import Any
  session = AgentSession(turn_detection=MultilingualModel()  # type: ignore[arg-type]
  )

Use

cast

or

# type: ignore

sparingly; prefer a proper interface import when possible. If you share the exact imports and the AgentSession signature from your environment (or the exact type error text and your type checker), I can give the precise import or cast to remove the warning.

from livekit.plugins.turn_detector.multilingual import MultilingualModel

This is the import statement

Thanks — with your import confirmed, the warning means the static type expected by

AgentSession.turn_detection

doesn't match the concrete class

MultilingualModel

. Quick fixes (pick one): 1) Use the SDK's TurnDetection type if available

from livekit.agents.types import TurnDetection
from livekit.plugins.turn_detector.multilingual import MultilingualModel

turn_detection: TurnDetection = MultilingualModel()
session = AgentSession(turn_detection=turn_detection, ...)

2) Cast to Any (safe, minimal change)

from typing import Any
from livekit.plugins.turn_detector.multilingual import MultilingualModel

session = AgentSession(turn_detection=cast(Any, MultilingualModel()), ...)
# or
session = AgentSession(turn_detection=MultilingualModel(),)  # type: ignore[arg-type]

3) Ensure package version and extras • Upgrade so type hints are present: pip install -U "livekit-agents[turn-detector]" • Confirm your installed version (share output of

pip show livekit-agents

if unsure). If you want, share: • the exact type-checker/IDE warning text and •

pip show livekit-agents

output and I’ll give the precise import or cast to clear the warning. Sources: Improving voice AI's turn detection with transformers | LiveKit turn detector plugin | LiveKit docs | Silero VAD plugin | LiveKit docs | Improved End-of-Turn Model Cuts Voice AI Interruptions 39%