This sounds less like a transcription/model issue and more like an audio pipeline problem.
Your Deepgram settings won’t affect the actual recording quality—they only impact transcription behavior. The distortion and low volume usually point to gain staging, input clipping, or encoding issues somewhere in the call pipeline (mic input, WebRTC stream, or how VAPI stores/normalizes the recording).
Given it happens at higher volumes, it’s likely clipping or compression during capture. I’d suggest checking the raw input levels and whether any automatic gain control or noise suppression is overcorrecting.
If you can share a sample or logs, I can help pinpoint exactly where it’s breaking
@esra