Turn Detector Error: "Turn detector does not support language english"

Server SDKs
1

The MultilingualModel turn detector reports it doesn’t support language “english” even though the STT provider is configured with language=“en”.

Environment & Versions

  • Python: 3.12 (or 3.11+ as minimum requirement)
  • livekit-agents: ^1.4
  • livekit-plugins-openai: ^1.4
  • livekit-plugins-turn-detector: ^1.4
  • livekit-api: ^1.1
  • STT Provider: OpenAI Whisper (whisper-1)
  • Turn Detector: MultilingualModel() (no parameters)

Error Logs -

2026-02-17 05:12:02.448 | INFO | logging | callHandlers:1762 | STT metrics
{
    "message": "STT metrics",
    "level": "INFO",
    "name": "livekit.agents",
    "model_name": "whisper-1",
    "model_provider": "api.openai.com",
    "audio_duration": 5.56,
    "session_id": "AJ_P6zq2755rWKK",
    "round_id": "7fd76d6e-3a30-4eb8-8a3c-385ae385d00e",
    "account_id": "90f046d7-4abf-4830-a512-d94801db3bf2",
    "pid": 500,
    "job_id": "AJ_P6zq2755rWKK",
    "room_id": "RM_tSFbsDYGHRQu",
    "timestamp": "2026-02-17T05:12:02.447572+00:00"
}

2026-02-17 05:12:02.448 | INFO | logging | callHandlers:1762 | Turn detector does not support language english
{
    "message": "Turn detector does not support language english",
    "level": "INFO",
    "name": "livekit.agents",
    "session_id": "AJ_P6zq2755rWKK",
    "round_id": "7fd76d6e-3a30-4eb8-8a3c-385ae385d00e",
    "account_id": "90f046d7-4abf-4830-a512-d94801db3bf2",
    "pid": 500,
    "job_id": "AJ_P6zq2755rWKK",
    "room_id": "RM_tSFbsDYGHRQu",
    "timestamp": "2026-02-17T05:12:02.448677+00:00"
}

Relevant Code

STT Provider Configuration:

def _setup_stt_providers(self) -> List[Any]:
    """Set up STT providers with fallback mechanism"""
    stt_providers = []

    # Primary STT provider: OpenAI Whisper
    if settings.OPENAI_API_KEY:
        stt_providers.append(
            openai.STT(
                api_key=settings.OPENAI_API_KEY,
                language="en",  # ISO 639-1 code
                model=DEFAULT_STT_MODEL,  # "whisper-1"
            )
        )
        logger.info("OpenAI STT configured as primary provider")

    # Fallback STT provider: Deepgram (if available and configured)
    if settings.DEEPGRAM_API_KEY:
        stt_providers.append(
            deepgram.STT(
                api_key=settings.DEEPGRAM_API_KEY, 
                model="nova-2", 
                language="en"  # ISO 639-1 code
            )
        )
        logger.info("Deepgram STT configured as fallback provider")

    return stt_providers

AgentSession Configuration:

from livekit.plugins.turn_detector.multilingual import MultilingualModel
from livekit.agents.voice import AgentSession

# ... STT setup code above ...

# Create agent session with the selected TTS, STT, and LLM
self.session = AgentSession[InterviewData](
    userdata=self.interview_data,
    stt=stt_instance,  # OpenAI STT with language="en"
    llm=llm_instance,
    tts=tts_instance,
    vad=self.vad,
    turn_detection=MultilingualModel(),  # No language parameter
    min_endpointing_delay=2.0,
    max_endpointing_delay=3.0,
    max_tool_steps=10,
)

I’m not sure what’s causing this issue. Any help would be appreciated!

2

The issue is that the MultilingualModel turn detector is receiving the language code "english" (the full language name) instead of the expected ISO 639-1 code "en" (the two-letter code).

3

The way i see it. The turn detector infers the language info from STT. So, isn’t this a livekit issue where the wrong language format is being passed?

4

How are you defining your fallback adapter? I mean, what do you do with the return value of _setup_stt_providers?

5

Putting the relevant code here -

async def start(self) -> None:
    """Start the interview session"""
    logger.info("Starting interview session")

    # Validate settings
    validate_required_settings()

    # Ensure interview data is initialized
    if not self.interview_data or not self.vad:
        raise ModuleError(
            "livekit_agent.interview.voice.session",
            "Interview session not properly initialized. Call initialize() first.",
        )

    # Get voice settings from the interview data
    voice_settings = self.interview_data.interview_data.get("voice_settings", {})

    # Set up providers with fallback mechanisms
    tts_providers = self._setup_tts_providers(voice_settings)
    stt_providers = self._setup_stt_providers()
    llm_providers = self._setup_llm_providers()

    # Create provider instances with fallback support
    tts_instance = self._create_provider_instance(tts_providers, "TTS")
    stt_instance = self._create_provider_instance(stt_providers, "STT")
    llm_instance = self._create_provider_instance(llm_providers, "LLM")

    # Create agent session with the selected TTS, STT, and LLM
    self.session = AgentSession[InterviewData](
        userdata=self.interview_data,
        stt=stt_instance,
        llm=llm_instance,
        tts=tts_instance,
        vad=self.vad,
        turn_detection=MultilingualModel(),
        min_endpointing_delay=2.0,
        max_endpointing_delay=3.0,
        max_tool_steps=10,
    )


def _setup_stt_providers(self) -> List[Any]:
    """Set up STT providers with fallback mechanism

    Returns:
        List of configured STT providers
    """
    stt_providers = []

    # Primary STT provider: OpenAI Whisper
    if settings.OPENAI_API_KEY:
        stt_providers.append(
            openai.STT(
                api_key=settings.OPENAI_API_KEY,
                language="en",
                model=DEFAULT_STT_MODEL,
            )
        )
        logger.info("OpenAI STT configured as primary provider")

    # Fallback STT provider: Deepgram (if available and configured)
    if settings.DEEPGRAM_API_KEY:
        stt_providers.append(
            deepgram.STT(
                api_key=settings.DEEPGRAM_API_KEY, model="nova-2", language="en"
            )
        )
        logger.info("Deepgram STT configured as fallback provider")
    elif not settings.DEEPGRAM_API_KEY:
        logger.warning(
            "Deepgram API key not configured - STT fallback limited to OpenAI only"
        )

    return stt_providers


def _create_provider_instance(
    self, providers: List[Any], provider_type: str
) -> Any:
    """Create a provider instance with fallback support

    Args:
        providers: List of configured providers
        provider_type: Type of provider (TTS, STT, LLM) for logging

    Returns:
        Provider instance with fallback if multiple providers available

    Raises:
        ModuleError: If no providers are configured
    """
    if len(providers) > 1:
        if provider_type == "STT":
            if not self.vad:
                raise ModuleError(
                    "livekit_agent.interview.voice.session",
                    "VAD must be initialized before creating STT provider instances",
                )

            wrapped_providers = [
                stt.StreamAdapter(stt=provider, vad=self.vad)
                for provider in providers
            ]
            instance = stt.FallbackAdapter(cast(list[stt.STT], wrapped_providers))
        elif provider_type == "TTS":
            instance = tts.FallbackAdapter(providers)
        else:  # LLM
            instance = llm.FallbackAdapter(providers)
        logger.info(
            f"{provider_type} fallback enabled with {len(providers)} providers"
        )
        return instance
    elif len(providers) == 1:
        logger.warning(
            f"{provider_type} fallback not available - only one provider configured"
        )
        if provider_type == "STT":
            if not self.vad:
                raise ModuleError(
                    "livekit_agent.interview.voice.session",
                    "VAD must be initialized before creating STT provider instance",
                )

            return stt.StreamAdapter(stt=providers[0], vad=self.vad)
        return providers[0]
    else:
        raise ModuleError(
            "livekit_agent.interview.voice.session",
            f"No {provider_type} providers configured",
        )

So,

_setup_stt_providers():

  • Creates raw STT provider instances (OpenAI, Deepgram)
  • Configures each with language=“en” and API keys
  • Returns a list of provider instances: [openai.STT(language=“en”), deepgram.STT(language=“en”)]

_create_provider_instance():

  • Takes the list of providers and wraps them for use

  • For STT specifically:
    – If multiple providers: wraps each in stt.StreamAdapter (adds VAD), then wraps all in stt.FallbackAdapter
    – If single provider: wraps only in stt.StreamAdapter

6

Thanks for the detailed code, it really helps.

It looks like this issue is specific to OpenAI’s ‘whisper-1’ model. If I change this to ‘gpt-4o-transcribe’ the issue goes away.

AI is saying the root cause is this line: agents/livekit-plugins/livekit-plugins-openai/livekit/plugins/openai/stt.py at main · livekit/agents · GitHub, which causes the plugin to report its language as “english” rather than “en”, but I’m unsure why a specific check is made in the codebase there.

I can also reproduce this with our starter agent, GitHub - livekit-examples/agent-starter-python: A complete voice AI starter for LiveKit Agents with Python., and making a single change: stt=openai.STT(model="whisper-1"),

Is the whisper-1 model a hard requirement? If not, I suggest using a different model.

Let me follow up with the team to see if this is a known issue

7

Thanks for checking this out!
I ended up moving to a different STT, so I’m good now.

thanks again!