We are trying to reduce the latency on our calls. We are using a custom LLM so some of the issues could come from that. But we are using 11Labs 2.5 Turbo so are unsure why there is so much latency from that as well. Here is our latency summary: { "id": "log-1373", "level": 30, "time": 1757002359842, "timestamp": 1757002359842000000, "severityText": "INFO", "severityNumber": 9, "body": "Turn latency summary", "attributes": { "category": "latency", "callId": "4be01233-b2d4-498f-81e6-60ddffc90a96", "orgId": "a306bc70-3edb-4b5d-9fa0-ec3a0f78d0cb", "totalLatencyMs": 8935, "transcriberLatencyMs": 369, "endpointingLatencyMs": 501, "knowledgeBase": null, "model": { "fetchSequenceLatencyMs": 1, "startTokenLatencyMs": 3898, "firstOutputTokenLatencyMs": 3900, "avgTokenLatencyMs": 77 }, "voice": { "fromTranscriberLatencyMs": 8566, "fromModelLatencyMs": 4140, "fromRequestLatencyMs": 240 } } } And the voice settings "voice": { "provider": "11labs", "voiceId": "OYTbf65OHHFELVut7v2H", "model": "eleven_flash_v2_5", "speed": 0.94, "stability": 0.6, "similarityBoost": 0.78 },

Thanks for sharing the detailed latency breakdown and your 11Labs voice settings. To help narrow down the cause, could you clarify if this high latency (e.g., ~8,935 ms total) is consistent across all calls, or does it vary depending on the call or time of day?

@User It is pretty consistent

To reduce ElevenLabs TTS latency, try increasing the

speed

parameter closer to the upper limit (e.g.,

1.2

) for faster speech synthesis. Here’s how you can adjust it in your voice settings:

json
"voice": {
  "provider": "11labs",
  "voiceId": "OYTbf65OHHFELVut7v2H",
  "model": "eleven_flash_v2_5",
  "speed": 1.2,
  "stability": 0.6,
  "similarityBoost": 0.78
}

This change can help decrease the time ElevenLabs takes to generate audio, improving overall call responsiveness. Source: - [Changelog March 2, 2025: ElevenLabs speed parameter enhancement](https://docs.vapi.ai/changelog/2025/3/2)

I don't think that is right. I think the speed is the speed of the voice, not the speech synthesis speed

"firstOutputTokenLatencyMs": 3900 Seems like an LLM issue

@A Igor Good idea. We just tried that and it does seem to be better. I am still curious why the "fromMOdelLatencyMs" on the voice was 4140 in this instance. Do you know what that is specifically?

From model latency is from LLM first token to voice first token So it includes the latency from the language model as well

Okay, thank you!

Hi Noah Johnson, For addressing latency issues with 11Labs, it's important to check if the Eleven Labs voice service has been disabled. There's a new call end reason

pipeline-error-eleven-labs-vapi-voice-disabled-by-owner

which indicates if the service is turned off by the owner. You might want to verify this setting. If this is not the case, for a more detailed investigation, please consider providing: - The Call ID from the problematic call - A timestamp when the issue occurred - A description of the expected vs. actual outcomes For more guidance, you can refer to the Vapi [issue reporting documentation](https://docs.vapi.ai/issue-reporting) to assist in structuring your issue report effectively. If this doesn’t resolve the issue, please let us know and we’ll be happy to step in and assist further.