This issue needs urgent attention. The agent ran perfectly on n8n Cloud, but after moving to a self hosted n8n instance and adding new ElevenLabs and OpenAI keys, it now freezes mid sentence during calls. We have ruled out all local and workflow issues: • n8n workflows respond instantly and return 200 OK from PowerShell tests • Tested on two separate internet connections (hotel WiFi and mobile hotspot) • Tried different Vapi agents, models, and latency modes (More Latency on, Auto Mode off) • Deleted and recreated ElevenLabs API key, reconnected cleanly • ElevenLabs usage logs show successful requests, no quota or auth errors • Call logs show ElevenLabs WebSocket opened then closing mid turn with no code • Stream drops only during mid speech, not at start or end • Issue replicates across devices and networks Everything points to a streaming or buffer issue between Vapi and ElevenLabs. Please check backend logs for recent calls (for example call ID 019a4a11 ae20 7446 8512 c024821a759c) and confirm the cause of these mid turn stream drops. It is midnight here in Vietnam. We will be back online in about 7 hours to continue testing. Also, Discord-only support is really poor for paying customers, i was limited on my characters unless i upgrade. This needs proper ticket escalation, i had previously written a detailed description of the issues. Social Brainstorm

Actionable Next Step: Test your voice agent directly in the [Vapi dashboard](https://dashboard.vapi.ai/) using the "Talk to Assistant" feature. This will help you determine if the freezing issue is caused by your self-hosted n8n server's network configuration or if it persists independently of your environment. If the agent works smoothly in the dashboard, review your server's outbound firewall and WebSocket settings to ensure ElevenLabs connections are not being interrupted. This quick isolation test will pinpoint whether the problem is with your server or the Vapi/ElevenLabs integration itself ([Debugging Voice Agents](https://docs.vapi.ai/debugging#test-in-dashboard)). Source: - [Debugging Voice Agents](https://docs.vapi.ai/debugging#test-in-dashboard)

No, I’ve already done all of those tests. I’m going to email the support team directly because I’ve been waiting too long for a proper response. This is the second message I’ve sent the first one was deleted, and now I’m being restricted by Discord’s character limit, which is why you haven’t seen the full list of tests and logs we’ve already completed.

I understand how frustrating that must feel after all the testing you’ve done. I can help you dig deeper into the root cause — this kind of mid-turn drop usually points to a WebSocket timeout or proxy interference on the self-hosted setup. Could you confirm if your n8n instance is running behind Nginx, Cloudflare, or any reverse proxy? Also, check if persistent connections and streaming are fully enabled in your environment. Once I know that, I can guide you through stabilizing the stream between Vapi and ElevenLabs. @socialbrainstorm

Hi Vercel, Thanks for getting back to me. Yes, we’ve already tested the full setup and confirmed there’s no proxy or timeout issue on our end. The n8n instance runs self-hosted with direct HTTPS access no Cloudflare, no Nginx reverse proxy, and no extra firewall layers. We’ve also emailed a detailed diagnostic report earlier today with full logs, test results, and PowerShell verification confirming the ElevenLabs key is valid and streaming is open. Please review that email for the full technical breakdown when you can it includes timestamps and sample WebSocket traces. Appreciate your help, James Woolford

Hi James, Thanks for the detailed update. I’m not part of ElevenLabs or n8n support, but as a specialist, I can help analyze your setup and pinpoint why the stream drops mid-turn. Based on what you’ve described, it still seems like a buffer or streaming handling issue between Vapi and ElevenLabs that could use a deeper dive. We can discuss your logs and WebSocket traces privately so I can guide you step by step to stabilize the connection. @socialbrainstorm

thank you, for your time. ive fixed the issue.