Hi everyone, I’m testing a French voice agent over phone calls and I’m having an audio capture issue: the LLM/STT often doesn’t hear the speaker well unless the caller speaks quite loudly. What we tested so far: • We removednoise_cancellationfromRoomInputOptions, and that improved things noticeably on desktop / local microphone input. • However, on smartphone calls through telephony, the problem is still there: the caller is sometimes only partially heard, or not heard well enough unless speaking loudly. • We are using: ◦ deepgram.STT(model="nova-3", language="fr") ◦ silero.VAD.load() ◦ turn_detection="vad" • We also checked Deepgram model options, but for French there doesn’t seem to be a dedicated phone-call model like the Englishnova-2-phonecall. • On the telephony side, we are also looking at codec ordering on the SIP trunk (G711U,G711A,G722,G729) to see whether that helps. My question is: What is the best way in LiveKit to improve how well the agent hears a caller on phone calls, especially when the incoming voice level is low? More specifically: • Are there recommended LiveKit settings for telephony audio input sensitivity? • Should we tune Silero VAD for this kind of issue, or is that mostly turn detection rather than capture? • Any known best practices for Telnyx + LiveKit + Deepgram when callers are not picked up reliably unless they speak loudly? Any guidance would be really appreciated.

Hi Christophe Chapiteau, ⚠️ This channel will be deleted soon ⚠️ Please use our new community at https://community.livekit.io for technical discussions and to post technical questions. The Slack community is not going away, but we are migrating technical questions to our new forum. Please see full details here.