Hi team, We’re facing an issue with our Vapi assistant where we’re seeing noticeable response delays during live calls. What’s happening: • The assistant sometimes pauses for several seconds before responding. • In some cases, it stays silent and then follows up with “Are you still there?”, which breaks the conversation flow. • This happens even after simple user responses like confirmations or yes/no answers. Current setup: • Prompt behavior is working as expected. • Temperature is set to 0.15 (tested lower values as well). • The issue seems to be related to latency, not response accuracy. What we’ve tried: • Prompt refinements and restructuring • Testing lower temperature values • Reviewing call logs and workflow logic • General debugging on our side Despite this, the delayed responses still persist in real-time calls. Could you please advise if there are any known causes, configurations, or best practices to reduce response latency during live calls? Happy to share more details if needed. Thanks!

To reduce response latency, try making your smart endpointing more aggressive and lowering the wait time before the assistant speaks. Update your

startSpeakingPlan

as follows:

json
{
  "startSpeakingPlan": {
    "waitSeconds": 0.1,
    "smartEndpointingPlan": {
      "provider": "livekit",
      "waitFunction": "200 + 1000 * x"
    }
  }
}

This should help your assistant respond faster to simple confirmations and yes/no answers. For more details, see the [voice pipeline configuration documentation](https://docs.vapi.ai/customization/voice-pipeline-configuration). Source: - [Voice pipeline configuration documentation](https://docs.vapi.ai/customization/voice-pipeline-configuration)

Thanks for the suggestion regarding making the startSpeakingPlan more aggressive to reduce response latency. We’ve already tried similar optimizations on our end and have updated the assistant configuration accordingly. Below is the current setup we are using after making these changes: { "idleMessageMaxSpokenCount": 1, "idleTimeoutSeconds": 5, "startSpeakingPlan": { "waitSeconds": 0.1, "transcriptionEndpointingPlan": { "onPunctuationSeconds": 0.1, "onNoPunctuationSeconds": 0.3, "onNumberSeconds": 0.3 }, "smartEndpointingPlan": { "provider": "livekit", "waitFunction": "2000 / (1 + exp(-10 * (x - 0.5)))" } }, "stopSpeakingPlan": { "numWords": 0, "voiceSeconds": 0.2 } } These changes did help slightly, but we’re still observing noticeable delays in real-time calls — especially after simple confirmations or yes/no responses, where the assistant sometimes pauses before speaking. Please let us know if: • there are additional parameters we should tune, or • a different endpointing strategy / waitFunction is recommended for such conversational use cases. Happy to test any further suggestions you may have. Best regards, Vaibhav Jain

Hi Vj! To help track down this issue, could you share: - The call ID - When exactly this happened (the timestamp) - What response you expected to get - What response you actually got instead This would really help us figure out what went wrong!

Hi Praveen, Thanks for getting back to us. We wanted to clarify that we’ve already shared all the requested details over email with the Vapi support team (support@vapi.ai) earlier, including multiple examples with call IDs, timestamps, and exact observed behavior. We were also in touch with Kyle over a call regarding the same. However, it’s been close to a month now without any follow-up or resolution, and we’re currently at the final stage of platform selection. Given the criticality of this decision, timely clarity from your side is extremely important for us.

I've analyzed the reported call latency issues for the three calls. Here's my comprehensive findings: Issue Summary Three calls experienced 4-16 second response delays between user speech completion and assistant response, causing incorrect "Are you still there?" triggers. Call IDs: •

019b2624-577c-7334-b05b-8b1cff191f1c

•

019b262b-532a-7778-bf69-aee5e90eabd5

•

019b262e-1501-788d-9cff-22cea616e251

Root Cause Analysis Based on Vapi's latency architecture and similar documented incidents, the delays likely stem from: 1. Endpointing Latency Spikes The default

onNoPunctuationSeconds

is 1500ms, which can cause perceivable delays while waiting to confirm the user finished speaking. 2. LLM Model Latency Model latency (time to first token) can spike to 1200-1700ms during peak load or provider issues. 3. Knowledge Base Retrieval Timeouts If KB was configured, the default 15-second timeout can cause silence-timeout triggers. 4. Compounding Effects The inconsistency of delays (4s, 7s, 8s, 13s, 16s) suggests multiple factors compounding at different points in the pipeline. Recommendations 1. **Optimize Endpointing**: Use aggressive wait function, reduce

onNoPunctuationSeconds

to 500-800ms 2. **Set KB Timeout**: Reduce from 15s default to < 5 seconds 3. **Monitor LLM Latency**: Set alerts for >1000ms spikes 4. **Configure Fallback Providers**: For both LLM and TTS 5. **Review Assistant Config**: Check endpointing settings, wait function, custom rules Next Steps To fully diagnose, we need to: 1. Review the assistant's configuration for endpointing settings 2. Check infrastructure metrics for Jan 27 around 1:22-1:33 PM 3. Verify Knowledge Base configuration and timeouts if applicable Let me know if you'd like me to dig deeper into any specific area or if you can share the assistant configuration for these calls.

Thanks a lot for the detailed analysis — really appreciate the quick and clear response. I’ll first try the suggested changes on our side (endpointing, KB timeout, etc.) and see how things improve. If the issue gets resolved, that’s great. Otherwise, if you’re available, we’d love to set up a quick call to go through this more deeply. I’ll also share the assistant configuration details with you as soon as possible.