Struggling With Bot Cutting Off Callers—Need Longer Pause Settings / Endpointing Advice
Hey Vapi team, hoping someone can offer insight or suggestions!
I’m running a voice assistant for my company, and I’m struggling to get the right balance between latency and patience when callers give complex responses—especially with addresses (e.g., “892 Southwest 14th Street, Miramar 33022”). The biggest problem is that Jamie often cuts off callers before they’re finished, even when I set the pause detection (“onNumberSeconds”, “onPunctuationSeconds”, “onNoPunctuationSeconds”) to the max (3 seconds).
What I’ve tried so far:
Set all endpointing parameters to 3 seconds (the max in the Vapi UI)
Tested with Deepgram Nova 2, Deepgram Nova 3, and OpenAI GPT-4o Transcribe models
Deepgram is much faster (low latency, ~750ms), but still cuts off callers on long responses
OpenAI GPT-4o is more patient, but the latency jumps to ~1500ms, which makes the experience feel sluggish and robotic
What I’d like to know:
Is there a way to push the endpointing/pause thresholds even higher (beyond 3 seconds) for certain questions or globally—maybe via API, advanced config, or feature request? (Even a few extra seconds would help with addresses.)
Is there any way to tune or override the startSpeakingPlan and stopSpeakingPlan beyond what’s available in the UI, to better handle “chunky” user responses (e.g., when callers pause between house number, street, and city)?
Has anyone found a workaround to keep latency low but make the bot “wait longer” before responding, especially for multi-part responses like addresses or emails?
Is there a recommended combo of model + settings that works best for real-world phone calls where users naturally pause or spell things out?
I want to deliver a natural, patient, human-like experience—but I keep hitting the wall of either cut-off responses (if latency is low) or robotic/slow responses (if patience is high).
Any advice, advanced tricks, or suggestions would be hugely appreciated!
I’m running a voice assistant for my company, and I’m struggling to get the right balance between latency and patience when callers give complex responses—especially with addresses (e.g., “892 Southwest 14th Street, Miramar 33022”). The biggest problem is that Jamie often cuts off callers before they’re finished, even when I set the pause detection (“onNumberSeconds”, “onPunctuationSeconds”, “onNoPunctuationSeconds”) to the max (3 seconds).
What I’ve tried so far:
Set all endpointing parameters to 3 seconds (the max in the Vapi UI)
Tested with Deepgram Nova 2, Deepgram Nova 3, and OpenAI GPT-4o Transcribe models
Deepgram is much faster (low latency, ~750ms), but still cuts off callers on long responses
OpenAI GPT-4o is more patient, but the latency jumps to ~1500ms, which makes the experience feel sluggish and robotic
What I’d like to know:
Is there a way to push the endpointing/pause thresholds even higher (beyond 3 seconds) for certain questions or globally—maybe via API, advanced config, or feature request? (Even a few extra seconds would help with addresses.)
Is there any way to tune or override the startSpeakingPlan and stopSpeakingPlan beyond what’s available in the UI, to better handle “chunky” user responses (e.g., when callers pause between house number, street, and city)?
Has anyone found a workaround to keep latency low but make the bot “wait longer” before responding, especially for multi-part responses like addresses or emails?
Is there a recommended combo of model + settings that works best for real-world phone calls where users naturally pause or spell things out?
I want to deliver a natural, patient, human-like experience—but I keep hitting the wall of either cut-off responses (if latency is low) or robotic/slow responses (if patience is high).
Any advice, advanced tricks, or suggestions would be hugely appreciated!
