hello

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.

ElevenLabs inference.STT() outputs wrong language script (Malayalam) when speaking English Hi, I'm experiencing hallucinations with ElevenLabs STT via the inference addon. When speaking clearly in English, the transcription comes back in Malayalam script (Indian language). Configuration:

stt=inference.STT(model="elevenlabs/scribe_v2_realtime", language="en")

SDK Version: livekit-agents 1.3.7 Example Debug Output:

transcript='ഹലോ കെന്യൂ ഹിയർ മി ഐ സെഡ്' is_final=False language='en'

The text

ഹലോ കെന്യൂ ഹിയർ മി ഐ സെഡ്

is Malayalam script that phonetically reads "hello can you hear me I said" - so ElevenLabs is hearing the English correctly but outputting it in the wrong script. Important: Deepgram via

inference.STT(model="deepgram/nova-3", language="en")

works perfectly with the exact same setup. The issue is specific to ElevenLabs through the inference layer. I saw issue #4255 mentioning a fix for how

scribe_v2_realtime

model name is sent to providers. Has this fix been released? I'm on 1.3.7 and still experiencing this issue.

Thanks — I see you’re reporting ElevenLabs

scribe_v2_realtime

returning Malayalam-script text even though you set

language="en"

and that Deepgram works fine. You’re also asking whether the fix from issue #4255 has been released, since you’re on livekit-agents 1.3.7. Summary and current status • The incorrect-script behavior is caused by how the

scribe_v2_realtime

model name was forwarded through the inference layer. • A fix for that handling (issue #4255) has been merged to the main branch, but it is not included in the released SDK version you’re using (livekit-agents 1.3.7). What you can do now • Continue using Deepgram (

inference.STT(model="deepgram/nova-3", language="en")

) — this works correctly for your setup. • Or use the ElevenLabs plugin directly (bypass the inference layer) so you control request formatting:

from livekit.plugins import elevenlabs
  stt = elevenlabs.STT(model="scribe_v2_realtime")

This may avoid the model-name forwarding issue. Next steps I can help with • If you’d like, I can: ◦ Notify you when the fix ships in the next release. ◦ Walk through switching to the direct ElevenLabs plugin in your code. ◦ Help test a patch if you can run from main/edge builds. Which of those would you prefer? Sources: ElevenLabs STT plugin guide | LiveKit docs | ElevenLabs STT | LiveKit docs | Deepgram STT | LiveKit docs | AssemblyAI STT | LiveKit Docs