Why at got-realtime when model speaking it hearing itself and produce transcripts and think that it is turn

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

gpt-realtime model and krisp does not help

I'm here to help you work through the self-hearing or transcript issues you’re experiencing with the GPT-realtime model and Krisp. Let’s try these steps together: For self-hearing issues: • Make sure

echoCancellation=true

is turned on in your audio capture settings. This really helps prevent echo and avoids those frustrating feedback loops. • Check that the AI agent isn’t subscribing to its own audio track. It should only be picking up audio tracks with a

MICROPHONE

source from other participants, which keeps things clear. For transcript issues: • Consider adding a dedicated STT plugin like Google Cloud STT alongside the GPT-realtime model. This can give you low-latency, streaming transcriptions since the GPT-realtime model’s transcription tends to be delayed and might not work well for real-time display. • Try adjusting turn detection settings, such as

silence_duration_ms

, to make the system more responsive and reduce any false triggers. If you’d like, I’m happy to walk you through implementing these solutions or answer any questions you have. Just let me know! Sources: AI voice agents | LiveKit Docs | Development Tools | livekit/livekit_composite | DeepWiki | Realtime model integrations | LiveKit Docs | Live Conversations with AI using ChatGPT and WebRTC | livekit-agents/livekit/agents/llm/realtime.py