Hello, I'm testing out vapi for some sample integration but I observed huge latences for models that happen irregulary for example: { "id": "log-23", "level": 30, "time": 1760222076266, "timestamp": 1760222076266000000, "severityText": "INFO", "severityNumber": 9, "body": "Deepgram partial transcript", "attributes": { "category": "transcriber", "callId": "0199d569-4178-7dd7-bf2d-dc87336942e2", "orgId": "", "provider": "deepgram", "channel": "assistant", "transcript": "Mechanic on Wheels", "confidence": 0.9588216, "numWords": 3, "duration": 1040, "latency": 2844 } } what can be the cause ? (it's just 3 words)

A common cause of irregular latency (even for short utterances) is the TTS or STT model loading on the first request. To reduce this, you can "warm up" your model at server startup by sending a dummy request, ensuring the model is loaded before real calls arrive.

js
// Example: Warm up Deepgram or TTS model at startup
await deepgramClient.transcription.preload({ text: "test", sampleRate: 16000 });

Try adding a model warm-up step in your server initialization to minimize first-call latency. For more, see the performance optimization tips in the [Vapi documentation](https://docs.vapi.ai/customization/custom-voices/custom-tts). Source: - [Custom TTS integration](https://docs.vapi.ai/customization/custom-voices/custom-tts)

I can help you look into that. Latency spikes like this can come from network delays, Deepgram model load time, or temporary queueing on Vapi’s side. I can help you analyze the logs to see if it’s model-related or connection-related. Could you share if the issue happens consistently with certain models or only randomly across calls? @JustAnotherAIUser

Thanks for responding. It happened few times on irregular basis. Worth mentioning it happens on the intial message from voice agent, so no conversation started yet

Thanks for the update. I can help you identify the cause, it’s likely from model cold start or audio stream setup before the first message. Could you share one full call log or timestamp where it occurred so I can review it?

Ive sent a PM, but also sleep time for me. Happy to chat tomorrow.