yeah I actually had this exact issue in a voicemail detection system I built and had to solve by processing interim transcriptions in a lower-level orchestration framework. I'd be surprised if you can fix this with Vapi because, as you identified, the issue is the VAD never gets to send the user stopped speaking event, and I can't think of a way to configure a Vapi pipeline to process buffered interim transcriptions and trigger a tool based on what's been accumulated.