speech_final marks silence, not the end of a thought. At 300ms endpointing, 22 percent of users were cut off mid-answer. Raise the threshold before you start tuning prompts.

Context: Postmortem by ji_ai on dev.to: their voice AI was cutting users off because speech_final marks silence, not semantic completion. With endpointing at 300ms, 22 percent of users got cut off mid-answer. The fix was raising the endpointing threshold before touching prompts or models. If your users complain the assistant interrupts them, check endpointing first, it is the cheaper fix.