Hi! For the last 2 month or so we are experiencing long delays before assistants response (about 3-4 seconds). Estimated delay on assistant is about 1 sec, but even after getting it down to 850ms the overall delay is still about 3-4 seconds. We do not have any external tool or callback involved. Could you please help us figure out where exactly the bottleneck is? Example call id: 1f60e2d6-3e74-461c-841b-bdd919de19fd

Try adjusting these VAPI configuration settings to reduce response delays:

// Update via API: PUT /assistant/:id
{
  "silenceTimeoutSeconds": [lower value],
  "responseDelaySeconds": [lower value],
  "llmRequestDelaySeconds": [lower value],
  "smartEndpointingEnabled": false, // For English transcriptions
  "startSpeakingPlan": {
    "waitSeconds": 0.8
  }
}

Also set Deepgram's endpointing to 300ms for faster response times. Source: - [Discord discussion on configuration adjustments](https://discord.com/channels/1211482211119796234/1332286268805087272)

in your call log you can see

[INFO]

logs which gives you times for each step, for example https://cdn.discordapp.com/attachments/1369289747704582204/1369290242976256190/image.png?ex=681b526a&is=681a00ea&hm=0c22c3be2472648addc3ec7b099b64dd20086d4f39b5ad23c492212fc9412675&

The assistant's delayed response is due to end pointing, which determines the duration to wait before recognizing the user has finished speaking. - End Pointing Plan Configuration: You need to adjust the Smart End Pointing plan provider to livekit and update the wait function within the Start Speaking Plan for English Transcription. - Wait Function Optimization: Iterative testing with potential users is necessary to determine the optimal wait function values, as suggested values may not be suitable for all scenarios.

startSpeakingPlan": {
  "waitSeconds": 0.4,
  "smartEndpointingPlan": {
          "provider": "livekit",
          "waitFunction": "50 + 200 * x"
        }
}

Hi all! Here is my INFO log example: [INFO] Turn latency: 1198ms (transcriber: 302ms, endpointing: 104ms, kb: N/A, model: 413ms, voice: 362ms) I made changes to start speaking plan as suggested and it did help to some extend. I am wondering- is there a way to speed it up by 300-400 ms still?

maybe you could using groq llama 3.3 or llama 4 scout What transcriber / tts models are you using?

name: "Test" voice: { "model": "eleven_flash_v2", "speed": 1.1, "style": 0.2, "voiceId": "JZ3e95uoTACVf6tXaaEi", "provider": "11labs", "stability": 0.1, "similarityBoost": 0.3, "fillerInjectionEnabled": false, "optimizeStreamingLatency": 3 } createdAt: "2025-05-07T08:01:11.229Z" updatedAt: "2025-05-07T08:49:38.853Z" model: "gpt-4o-mini" role: "system" content: provider: "openai" temperature: 0.5 emotionRecognitionEnabled: true firstMessage: "Hi there! you’re talking to Sam! How are you doing today?" endCallFunctionEnabled: true endCallMessage: "" transcriber: { "model": "nova-3", "language": "en", "numerals": false, "provider": "deepgram", "confidenceThreshold": 0.4 } clientMessages: [ "conversation-update", "function-call", "hang", "model-output", "speech-update", "status-update", "transfer-update", "transcript", "tool-calls", "user-interrupted", "voice-input", "workflow.node.started" ] serverMessages: [ "end-of-call-report" ] hipaaEnabled: false maxDurationSeconds: 14089 backgroundSound: "off" backchannelingEnabled: false analysisPlan: { "structuredDataPrompt": "" } backgroundDenoisingEnabled: false startSpeakingPlan: { "waitSeconds": 0.4, "transcriptionEndpointingPlan": { "onPunctuationSeconds": 0.3, "onNoPunctuationSeconds": 1 }, "smartEndpointingEnabled": "livekit", "smartEndpointingPlan": { "provider": "livekit", "waitFunction": "50 + 200 * x" } } stopSpeakingPlan{ "numWords": 3 } compliancePlan{ "hipaaEnabled": false, "pciEnabled": false }

model: "gpt-4o-mini"

is as slow as 4o I believe, try 4.1 mini or even nano

To help track down this issue, could you share: - The call ID This would really help us figure out what went wrong!