https://livekit.io logo
Join Slack
Powered by
Livekit Interference addon questions
# ask-ai
l
Livekit Interference addon questions
t
Thanks for reaching out, G. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.
Thanks — I see you’re asking about the LiveKit Inference addon. I can help with that. Which of these would you like details on? • Pricing and billing for Inference • Setup and configuration for Agents using Inference • Which models are supported and how to select one • How Inference differs from using plugins (your own provider keys) • Troubleshooting latency, routing, or integration issues Or tell me your specific question/use case and I’ll provide targeted steps. Sources: Azure AI Services and LiveKit | LiveKit Docs | An open source stack for real-time multimodal AI | Introducing LiveKit Inference: A unified model interface for voice AI | OpenAI and LiveKit | LiveKit docs | Google Gemini LLM | LiveKit Docs
l
Questions for LiveKit Developer: Lithuanian STT Model Support Context We're building a B2B voice agent platform using LiveKit Agents SDK with LiveKit Inference. We need Lithuanian (lt) speech-to-text support. Our TTS and LLM work perfectly - only STT is failing. Question 1: What is the correct model ID format for ElevenLabs Scribe STT?
Copy code
What we tried:# backend/agent/main.py (line 185)
STT_MODEL = "elevenlabs/scribe-v2-realtime:lt"
The exact error:livekit.agents._exceptions.APIError: LiveKit STT returned error: 
{
  "type": "error",
  "session_id": "beef5948-a94a-4bfd-903d-57fc5f809585",
  "data": "Failed to create provider instance for session beef5948-a94a-4bfd-903d-57fc5f809585: model not supported by provider: scribe-v2-realtime",
  "code": 2004
}
1. *Questions:*Is 
elevenlabs
a valid provider for STT in LiveKit Inference? 2. What is the correct model ID for ElevenLabs Scribe V2 Realtime? 3. Does LiveKit Inference support ElevenLabs STT at all? Question 2: What STT providers support Lithuanian in LiveKit Inference?
Copy code
What we know works (English):STT_MODEL = "deepgram/nova-3:en"  # This worked perfectly
Our understanding of model format:<provider>/<model>:<language>
Working models we use:LLM_MODEL = "openai/gpt-4o-mini"                              # Works
TTS_MODEL = "cartesia/sonic-3:f786b574-daa5-4673-aa0c-cbe3e8534c02"  # Works
1. *Questions:*What STT providers support Lithuanian (
lt
) in LiveKit Inference? 2. Does 
deepgram/nova-2:lt
or 
deepgram/nova-3:lt
work? 3. What is the complete list of available STT models/providers? Question 3: How do we configure AgentSession for Lithuanian STT?
Copy code
Our current AgentSession configuration:# backend/agent/main.py (lines 356-364)
session = AgentSession(
    llm=inference.LLM.with_groq(model=config.get("llm_model", LLM_MODEL)),
    stt=inference.STT(model=config.get("stt_model", STT_MODEL)),
    tts=inference.TTS(model=config.get("tts_model", TTS_MODEL)),
    vad=silero.VAD.load(
        min_speech_duration=0.1,
        min_silence_duration=0.3,
    ),
    turn_detector=turn_detector,
)
Log showing our configuration was received:Creating AgentSession with STT=elevenlabs/scribe-v2-realtime:lt, LLM=openai/gpt-4o-mini, TTS=cartesia/sonic-3:f786b574-daa5-4673-aa0c-cbe3e8534c02
1. *Questions:*Is there a different API for specifying language separate from model? 2. Should we use 
inference.STT(model="...", language="lt")
instead? 3. Do we need to configure anything else for multi-language support? Question 4: Fallback - Does Deepgram support Lithuanian?
Copy code
If ElevenLabs is not available, what about:STT_MODEL = "deepgram/nova-2:lt"
# or
STT_MODEL = "deepgram/nova-3:lt"
1. *Questions:*Does Deepgram Nova-2 or Nova-3 support Lithuanian through LiveKit Inference? 2. What other providers support Lithuanian STT? Our Environment • LiveKit Agents SDK: Latest (from pip) • Deployment: Railway (Docker) • Python: 3.11 • LiveKit Cloud: Yes (production) What Works Perfectly • TTS: 
cartesia/sonic-3:...
• LLM: 
openai/gpt-4o-mini
• VAD: Silero • Turn Detector: MultilingualModel What Fails • STT: 
elevenlabs/scribe-v2-realtime:lt
→ "model not supported by provider" Summary Question 1. What is the correct STT model string to use for Lithuanian speech recognition with LiveKit Inference? Please provide:Exact model string format 2. Available providers that support Lithuanian 3. Any additional configuration needed
t
Thanks — I’m responding to your questions about Lithuanian STT (you noted ElevenLabs Scribe failing and Deepgram working in English). Short answers up front • LiveKit Inference does not provide a supported ElevenLabs STT model. ElevenLabs STT is available only via plugins, not via the managed Inference gateway. • LiveKit Inference does not expose any STT model that supports Lithuanian (
lt
). There is no valid Inference model string you can use for 
lt
. • Recommended path: use the Deepgram plugin (your Deepgram account/key) with 
nova-2
and specify 
language="lt"
(or 
language="multi"
for auto-detection). That will give you Lithuanian STT. Details and action steps 1) ElevenLabs Scribe STT • Is 
elevenlabs
a valid STT provider in Inference? No — ElevenLabs STT is not supported via LiveKit Inference. • Correct model ID format for ElevenLabs Scribe realtime? The plugin uses 
elevenlabs/scribe_v2_realtime
(underscores), but this is for the ElevenLabs plugin only and will not work through Inference. • Conclusion: your error ("model not supported by provider: scribe-v2-realtime") happens because the Inference gateway does not enable that STT provider/model. 2) Lithuanian support / available STT providers • LiveKit Inference: no STT model is exposed that supports 
lt
. • Deepgram: Deepgram Nova-2 supports Lithuanian in Deepgram’s API, but Inference may not expose 
lt
for managed Nova-2. To guarantee 
lt
, use the Deepgram plugin with your account. • Deepgram Nova-3 does not list 
lt
. • Summary of practical options: use Deepgram via plugin for Lithuanian; other Inference providers do not provide 
lt
. 3) AgentSession configuration for Lithuanian • You can set language separately using the STT constructor: ◦ Example (plugin approach):
Copy code
from livekit.plugins import deepgram

    session = AgentSession(
      stt=deepgram.STT(model="nova-2", language="lt"),
      llm=inference.LLM(model="openai/gpt-4o-mini"),
      tts=inference.TTS(model="cartesia/sonic-3", voice="f786b574-daa5-4673-aa0c-cbe3e8534c02"),
      ...
    )
• Using 
inference.STT(model="deepgram/nova-2", language="lt")
through Inference may fail if the gateway blocks `lt`; plugin is reliable. • For mixed-language audio, 
language="multi"
is recommended with Deepgram. 4) Fallback and recommended solution • Use Deepgram Nova-2 via the Deepgram plugin and your Deepgram API key for reliable Lithuanian STT. • Requirements for plugin: ◦ Add/install the Deepgram plugin for livekit-agents. ◦ Provide your 
DEEPGRAM_API_KEY
(plugin uses your account/billing). • If you prefer to remain on LiveKit Inference, you’ll need to wait until LiveKit adds 
lt
support or requests that LiveKit enable Deepgram 
lt
for Inference. If you want, I can: • Provide the exact plugin install command and minimal code snippet for your Docker/Railway setup. • Show a quick test script to validate Deepgram 
lt
transcription in your environment. Which would help you next? Sources: Speech-to-text (STT) models | LiveKit Docs | Build voice, video, and physical AI | LiveKit | ElevenLabs STT plugin guide | LiveKit docs | ElevenLabs STT | LiveKit docs | Deepgram STT | LiveKit docs
l
You say livekite inferene does not support scribe v2 but in docs you say otherwise : https://docs.livekit.io/agents/models/stt/inference/elevenlabs/
t
Thanks — I see the discrepancy and you’re right to point to the docs. I’ll clarify and help you get unstuck. What the docs say • LiveKit Inference does support ElevenLabs Scribe V2 Realtime. The documented model ID is: ◦ elevenlabs/scribe_v2_realtime • Scribe V2 Realtime supports multilingual transcription (automatic language detection) and can be used via Inference with a language tag, e.g.: ◦ 
elevenlabs/scribe_v2_realtime:lt
Why you saw the error • The error "model not supported by provider: scribe-v2-realtime" happens when the model identifier sent to the gateway doesn't match the supported name. Common causes: ◦ Using a dash instead of an underscore (
scribe-v2-realtime
vs 
scribe_v2_realtime
) ◦ Outdated SDK version that doesn’t include the provider/model mapping ◦ Typo or incorrect string format (provider/model:language) What to do next — quick checklist to fix 1. Use the documented model string with underscores: ◦ 
STT_MODEL = "elevenlabs/scribe_v2_realtime:lt"
1. Ensure your SDK is recent enough (use Python SDK v1.2.13+ or Node SDK v1.0.7+). 2. Specify STT in AgentSession either as a string or with `inference.STT`: ◦ As string:
Copy code
session = AgentSession(stt="elevenlabs/scribe_v2_realtime:lt", ...)
• Or explicitly:
Copy code
session = AgentSession(
       stt=inference.STT(model="elevenlabs/scribe_v2_realtime", language="lt"),
       ...
     )
1. Redeploy after updating code and SDK, then retry. If that still fails • Share the exact SDK version you’re using and the full error payload (you already shared one). I’ll help verify whether the gateway is rejecting the request or if a different configuration issue remains. Sources: ElevenLabs STT | LiveKit docs | Speech-to-text (STT) models | LiveKit Docs | Build voice, video, and physical AI | LiveKit | Models | LiveKit Docs
l
https://docs.livekit.io/agents/models/stt/inference/elevenlabs/ Questions for LiveKit Developer: Lithuanian STT Model Support Context We're building a B2B voice agent platform using LiveKit Agents SDK with LiveKit Inference. We need Lithuanian (lt) speech-to-text support. Our TTS and LLM work perfectly - only STT is failing. Question 1: What is the correct model ID format for ElevenLabs Scribe STT?
Copy code
What we tried:# backend/agent/main.py (line 185)
STT_MODEL = "elevenlabs/scribe-v2-realtime:lt"
The exact error:livekit.agents._exceptions.APIError: LiveKit STT returned error: 
{
  "type": "error",
  "session_id": "beef5948-a94a-4bfd-903d-57fc5f809585",
  "data": "Failed to create provider instance for session beef5948-a94a-4bfd-903d-57fc5f809585: model not supported by provider: scribe-v2-realtime",
  "code": 2004
}
1. *Questions:*Is 
elevenlabs
a valid provider for STT in LiveKit Inference? 2. What is the correct model ID for ElevenLabs Scribe V2 Realtime? 3. Does LiveKit Inference support ElevenLabs STT at all? Question 2: What STT providers support Lithuanian in LiveKit Inference?
Copy code
What we know works (English):STT_MODEL = "deepgram/nova-3:en"  # This worked perfectly
Our understanding of model format:<provider>/<model>:<language>
Working models we use:LLM_MODEL = "openai/gpt-4o-mini"                              # Works
TTS_MODEL = "cartesia/sonic-3:f786b574-daa5-4673-aa0c-cbe3e8534c02"  # Works
1. *Questions:*What STT providers support Lithuanian (
lt
) in LiveKit Inference? 2. Does 
deepgram/nova-2:lt
or 
deepgram/nova-3:lt
work? 3. What is the complete list of available STT models/providers? Question 3: How do we configure AgentSession for Lithuanian STT?
Copy code
Our current AgentSession configuration:# backend/agent/main.py (lines 356-364)
session = AgentSession(
    llm=inference.LLM.with_groq(model=config.get("llm_model", LLM_MODEL)),
    stt=inference.STT(model=config.get("stt_model", STT_MODEL)),
    tts=inference.TTS(model=config.get("tts_model", TTS_MODEL)),
    vad=silero.VAD.load(
        min_speech_duration=0.1,
        min_silence_duration=0.3,
    ),
    turn_detector=turn_detector,
)
Log showing our configuration was received:Creating AgentSession with STT=elevenlabs/scribe-v2-realtime:lt, LLM=openai/gpt-4o-mini, TTS=cartesia/sonic-3:f786b574-daa5-4673-aa0c-cbe3e8534c02
1. *Questions:*Is there a different API for specifying language separate from model? 2. Should we use 
inference.STT(model="...", language="lt")
instead? 3. Do we need to configure anything else for multi-language support? Question 4: Fallback - Does Deepgram support Lithuanian?
Copy code
If ElevenLabs is not available, what about:STT_MODEL = "deepgram/nova-2:lt"
# or
STT_MODEL = "deepgram/nova-3:lt"
1. *Questions:*Does Deepgram Nova-2 or Nova-3 support Lithuanian through LiveKit Inference? 2. What other providers support Lithuanian STT? Our Environment • LiveKit Agents SDK: Latest (from pip) • Deployment: Railway (Docker) • Python: 3.11 • LiveKit Cloud: Yes (production) What Works Perfectly • TTS: 
cartesia/sonic-3:...
• LLM: 
openai/gpt-4o-mini
• VAD: Silero • Turn Detector: MultilingualModel What Fails • STT: 
elevenlabs/scribe-v2-realtime:lt
→ "model not supported by provider" Summary Question 1. What is the correct STT model string to use for Lithuanian speech recognition with LiveKit Inference? Please provide:Exact model string format 2. Available providers that support Lithuanian 3. Any additional configuration needed LiveKit Assistant [4:41 PM] Thanks — I’m responding to your questions about Lithuanian STT (you noted ElevenLabs Scribe failing and Deepgram working in English). Short answers up front • LiveKit Inference does not provide a supported ElevenLabs STT model. ElevenLabs STT is available only via plugins, not via the managed Inference gateway. • LiveKit Inference does not expose any STT model that supports Lithuanian (
lt
). There is no valid Inference model string you can use for 
lt
. • Recommended path: use the Deepgram plugin (your Deepgram account/key) with 
nova-2
and specify 
language="lt"
(or 
language="multi"
for auto-detection). That will give you Lithuanian STT. Details and action steps 1) ElevenLabs Scribe STT • Is 
elevenlabs
a valid STT provider in Inference? No — ElevenLabs STT is not supported via LiveKit Inference. • Correct model ID format for ElevenLabs Scribe realtime? The plugin uses 
elevenlabs/scribe_v2_realtime
(underscores), but this is for the ElevenLabs plugin only and will not work through Inference. • Conclusion: your error ("model not supported by provider: scribe-v2-realtime") happens because the Inference gateway does not enable that STT provider/model. 2) Lithuanian support / available STT providers • LiveKit Inference: no STT model is exposed that supports 
lt
. • Deepgram: Deepgram Nova-2 supports Lithuanian in Deepgram’s API, but Inference may not expose 
lt
for managed Nova-2. To guarantee 
lt
, use the Deepgram plugin with your account. • Deepgram Nova-3 does not list 
lt
. • Summary of practical options: use Deepgram via plugin for Lithuanian; other Inference providers do not provide 
lt
. 3) AgentSession configuration for Lithuanian • You can set language separately using the STT constructor: ◦ Example (plugin approach):
Copy code
from livekit.plugins import deepgram

    session = AgentSession(
      stt=deepgram.STT(model="nova-2", language="lt"),
      llm=inference.LLM(model="openai/gpt-4o-mini"),
      tts=inference.TTS(model="cartesia/sonic-3", voice="f786b574-daa5-4673-aa0c-cbe3e8534c02"),
      ...
    )
• Using 
inference.STT(model="deepgram/nova-2", language="lt")
through Inference may fail if the gateway blocks `lt`; plugin is reliable. • For mixed-language audio, 
language="multi"
is recommended with Deepgram. 4) Fallback and recommended solution • Use Deepgram Nova-2 via the Deepgram plugin and your Deepgram API key for reliable Lithuanian STT. • Requirements for plugin: ◦ Add/install the Deepgram plugin for livekit-agents. ◦ Provide your 
DEEPGRAM_API_KEY
(plugin uses your account/billing). • If you prefer to remain on LiveKit Inference, you’ll need to wait until LiveKit adds 
lt
support or requests that LiveKit enable Deepgram 
lt
for Inference. If you want, I can: • Provide the exact plugin install command and minimal code snippet for your Docker/Railway setup. • Show a quick test script to validate Deepgram 
lt
transcription in your environment. Which would help you next?
t
I don't have the answer you're looking for. You could also try asking your question: • in one of the other Slack channels or • to https://deepwiki.com/livekit/livekit_composite which is trained on all LiveKit source code If you find the answer, please post it here to help others!
8 Views
Open in Slack
PreviousNext
Powered by