I want to use only stt to capture transcripts for both hindi and English what is the best approach

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

I want to use deep gram

Great — since you want to use Deepgram, here’s a concise recommended approach to capture Hindi and English transcripts in LiveKit. 1) Use a multilingual Deepgram model • Choose Deepgram Nova-3 and set

language="multi"

so the model auto-detects Hindi and English in the same stream. 2) Configure LiveKit Inference STT • Create an

inference.STT

instance pointing to Deepgram with

model="deepgram/nova-3"

and

language="multi"

. • Supply your Deepgram API key in the inference configuration (ensure it’s stored securely). 3) Audio routing & encoding • Send mixed mono or stereo audio with a sample rate supported by Deepgram (typically 16 kHz+). • For best results, send clean audio channels per participant when possible (one stream per speaker) so diarization and accuracy improve. 4) Region & latency • Host your inference agent close to India (LiveKit Cloud Mumbai / ap-south) to reduce latency for Hindi/English use cases. 5) Optional accuracy improvements • Provide optional custom vocabulary or phrase hints for common domain terms if Deepgram supports it for your plan. • If you need guaranteed higher accuracy for one language, you can route that participant/session to

language="hi-IN"

or

language="en-IN"

instead of

multi

. 6) Monitoring & fallback • Monitor transcript confidence scores and handle low-confidence segments (e.g., flag for human review or reprocess with language-specific model). If you’d like, I can: • Provide the specific LiveKit

inference.STT

configuration snippet for Deepgram, or • Help with the audio encoding settings you’re using (sample rate, channels) to ensure compatibility. Which would you prefer?, with interactive elements