Our VAPI implementation for a scientific conversational interview in German has the problem that the model does not always appropriately identify when to wait (e.g. when the user is thinking) or when to start speaking. Livekit Smart Endpointing might perhaps help. But the VAPI implementation is currently only available in English also though Livekit has recently been updated to support multi-language support. However, this the new Livekit version seems currently not supported in VAPI. Do you have any other ideas to address this issue? (We did play around with that wait time variable)

Disable Smart Endpointing English-only, switch to transcription-based endpointing with tuned

waitSeconds

and punctuation settings, or use a custom orchestrator with LiveKit or AssemblyAI for full multilingual control over turn-taking and pause detection..