Hello, I have a question regarding call latency in my recordings (see attached screenshot). In the Latency Summary section, it shows me Average Turn Latency 1292ms. However, when listening to a real recording, the actual response latency (from the end of the user's call to the start of the AI's response) is subjectively close to 3 seconds. My questions are: Why is there such a difference between the measured value of 1292ms and the actual real-time latency that I hear? How can I effectively reduce the overall response latency? Which of the components (Voice, LLM, Transcriber, Endpointing) have the greatest potential for optimization? Context and Goal: For comparison: when testing voice models (e.g. via Google AI Studio with Gemini Flash Live), latency is often less than one second and transcription is virtually flawless. I would like to achieve similar speed and fluency within VAPI. Do you have any specific recommendations for setup, model selection, or configuration that would help me minimize this latency and achieve near-instantaneous and smooth conversation? https://cdn.discordapp.com/attachments/1438460648982843472/1438460649473839104/Snimka_obrazovky_2025-11-13_o_10.28.10.png?ex=694277e0&is=69412660&hm=17027d0f822c99d8c7b6bf1ce386dd3c02463a3ab90c3f7cef7cc85a47bc76a5&

To reduce perceived response latency, try lowering the

waitSeconds

parameter in your voice pipeline configuration for faster assistant replies. For example:

json
{
  "startSpeakingPlan": {
    "waitSeconds": 0.2
  }
}

Set

waitSeconds

to 0.2 for near-instant responses—this is especially effective for real-time or gaming use cases (see documentation for more details). Source: - [Voice Pipeline Configuration Documentation](https://docs.vapi.ai/customization/voice-pipeline-configuration)

I currently have it set to 0. What problems can this cause in practice?

Hey! 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!