It is showing total latency 1665. And as you can see from the visual that after the latency summary there is an extra couple second latency and I cannot figure out what that is. Does anyone know why i am getting extra latency after the summary is completed? I look through the logs and dont see anything out of the ordinary? https://cdn.discordapp.com/attachments/1458511159698391244/1458511159954112523/lately55.PNG?ex=695fe7de&is=695e965e&hm=d08d6c2f60eebc6db8b8ecdb2ec7b47447f598e3f5ba5563fec64bf5dd89e7f9&

To pinpoint the source of extra latency, use the detailed call metrics now available in your call artifacts. Check the

Call.artifact.performanceMetrics

fields, especially

turnLatencies

and

turnLatencyAverage

, to see if a specific conversation turn or component (model, voice, transcriber, endpointing) is responsible for the delay:

json
{
  "turnLatencies": [
    {
      "modelLatency": ...,
      "voiceLatency": ...,
      "transcriberLatency": ...,
      "endpointingLatency": ...,
      "turnLatency": ...
    }
  ]
}

Focus on any outlier values after the summary event to identify the bottleneck. According to the August 9, 2025 changelog, these metrics are designed to help you debug such issues with greater detail (see documentation). Source: - [Changelog: August 9, 2025 Call Metrics & Artifact Improvements](https://docs.vapi.ai/changelog/2025/8/9)

Hi ! Colehaa, Check the per-turn breakdown in the call artifact at Call.artifact.performanceMetrics (see changelog: https://docs.vapi.ai/changelog/2025/8/9). Steps to pinpoint the extra seconds after the summary: 1. Inspect the specific turn in Call.artifact.performanceMetrics.turnLatencies that corresponds to the summary turn. 2. Compare fields: modelLatency, voiceLatency, transcriberLatency, endpointingLatency, turnLatency. The extra seconds will show up as an outlier in one of those fields. 3. If modelLatency or voiceLatency = 0 but turnLatency > 0, consider caching or pre-generated audio (responses may be cached; see related threads). If model/voice are nonzero, those are likely the source. 4. If endpointingLatency is large (\~onNoPunctuationSeconds), verify endpointing rules and customEndpointingRules regex/timeout—a mismatched rule or fallback timeout can add \~1.5s. 5. If none of the component fields explain the gap: - Check the end-of-call webhook and artifact updates for any post-processing steps (analysis, structuredOutput extraction) that run after the summary. - Look for cached-tool calls, tool execution times, or knowledge-base requests in the call artifact/analysis that could run after the model output. - Review webhook/hang events in the server-message/webhook logs for any "hang" or "speech-update" messages around that timestamp.

Message marked as helpful by @Kizzz! 🎉