https://vapi.ai/ logo
Join Discord
Powered by
Really bad latency
# support
s
Why is the latency so bad on my assistant? It's not usually like this. call ID: 019e54d8-c8df-7aad-bce1-96e941b0994b
c
Hi there, Thank you for your message. Our team is currently out of the office. We operate Monday through Friday, from 9:00 AM to 8:00 PM Pacific Standard Time (PST). We’ll get back to you as soon as possible during our normal business hours. If your message is urgent, please mark it accordingly or include “URGENT” in the subject line, and we’ll do our best to respond promptly. Warm regards, Vapi Customer Support Team
Hi Saien, I'm currently investigating this issue. Thank you for your patience. I’ll get back to you once I identify the root cause. Kind regards, Akash
Hi, After reviewing your call ID, I found that the latency during this call was caused by a temporary slowdown on elevenLabs and deepgram’s end, as both providers were experiencing slower than usual response times at that moment. this was not related to your configuration. That said, there’s one change that should help improve latency going forward. you’re currently using eleven_turbo_v2, which is an older ElevenLabs model. switching to eleven_flash_v2_5 should noticeably reduce response times on future calls. Kind regards, Akash
s
Thanks Akash, can you also please check this call? 019e5a64-3427-799e-9a1b-92e85756aa1c This was with the flash 2.5 model I believe
c
Hi Saien, Sure, I’ll review the call logs and get back to you after checking them. Thanks for your patience. Kind regards, Akash
Hey, Apologies for the delayed response. This call was using OpenAI GPT-4.1. Could you please verify the model on your end and try again? Kind regards, Akash
s
Yes I used gpt 4.1
c
Hey, You mentioned the Flash 2.5 model earlier. Could you please provide a recent Call ID where you experienced the latency issue so that I can investigate it further?
s
I’m using eleven labs flash 2.5 TTS model but the LLM for the voice agent is open AI gpt 4.1
This is the call you can review
c
this one: 019e5a64-3427-799e-9a1b-92e85756aa1c ?
s
yes
Here is a more recent call: 019e7a27-7aba-7000-950a-854377d7fbf4
Feedback from my client: There is a significant delay, between me answering the call and the agent starting the conversion. Plus it was too slow in some of the pauses between the sections.
The LLM is GPT 4.1, the TTS model is Eleven Labs eleven flash v2.5 and the STT is Deepgram nova 3
c
hey just looked into the logs, the mid call pauses are from your answerFAQs knowledge base lookups. each FAQ answer needs a retrieval step before the agent can reply, which added ~3.7s on that turn. The delay right after pickup is voicemail detection running before the agent speaks, which added ~1.9s on the call.
trim your FAQ knowledge base to fewer/smaller docs (Assistant -> Tools / Knowledge Base) to speed up answers, and disable voicemail detection if it's not essential for this campaign (Assistant -> Advanced -> Voicemail Detection). Your normal turns and voice model are already good, so these two changes should clear up the lag.
s
This is confusing for me because the knowledge base is literally one PDF file and it’s 2 or 3 pages long
So how can that be causing such a delay? Also disabling voicemail detection doesn’t sound like a viable solution because voicemail detection is a necessary feature of the workflow. Should I be using another voicemail detection method or updating the settings of the voicemail detection?
here is a more recent call with the issue still persisting: 019e7e3d-745c-7001-9e09-a13b90f5fe58 On this call, I removed the answerFAQs tool from the assistant entirely and the latency issue still persisted so that doesn't seem to be the issue: 019e7e3f-fc28-7000-b1b2-0fecdfe54445
I think i found the root cause of the issue: it was telnyx Check out this recent call with a Twilio phone number: 019e8241-c689-7002-8e97-fda5385b3e59 The latency issue is gone but for all the telnyx calls, there was a bad latency issue. Any idea why?
c
Hi Saien, Thanks for your patience. compared both calls in detail, it's not Telnyx: • latency was identical : Telnyx 019e7e3d-745c-7001-9e09-a13b90f5fe58 = 20/49 ms Twilio 019e8241-c689-7002-8e97-fda5385b3e59 = 20/50 ms • The slow turn on the Telnyx call was an LLM failover: GPT-4.1's primary endpoint was slow that turn, so it retried on a backup region with a cold cache, adding ~1.3 s. The Twilio call just didn't hit a failover this time, same model, same pipeline best fix for now would be try a lower latency model like gpt-4.1-mini or gpt-4o and test again. Kind Regards, Akash
s
Hi Akash, I respectfully disagree. I tried a variety of models with telnyx (basically every open AI model as well gemini and claude) and all of those calls had high latency. As soon as I switched over to twilio, I never experienced latency on a single call. Telnyx was definitely the issue
The reason why I chose telnyx initially was because they have native UK call routing unlike Twilio. But I believe if I use SIP trunking on Twilio then I can do UK call routing, is that correct?
c
Yes, you can do UK routing on Twilio. It actually offers UK numbers natively, so you don't strictly need SIP trunking, you can provision or import a Twilio UK number directly, or use a Twilio SIP trunk if you prefer. Both work for the UK. SIP trunking setup if you go that route: https://docs.vapi.ai/advanced/sip/sip-trunk
s
Twilio only has 3 processing regions: US1 (United States), IE1 (Ireland), and AU1 (Australia). There is no UK-specific region. The closest is IE1 (Ireland). And by default, all calls and messages (regardless of where the phone number is from) are routed to US1.
So that’s why I want to use SIP trunking because I believe that could provide me with more options for routing the call closer to the UK
Worst case I’ll just route to Ireland
c
If Twilio is working correctly, then you can definitely route the calls to Ireland. Kind Regards, Akash
PreviousNext
Powered by