https://vapi.ai/ logo
Join Discord
Powered by
Major production issue.
# support
k
I’m experiencing multiple issues with our AI voice agents that started a few days ago. Our setup was previously working reliably with the same configuration, but performance has recently degraded and is impacting our business. Below are the specific issues and questions for clarity: 1. Abrupt Speech Stopping - The agent starts speaking but stops abruptly mid-response. Example Call ID: 019da136-e982-7001-966e-16bcbe1e8c48 (timestamp: 0:16) - What could be causing this interruption? - Has anything changed recently in streaming, buffering, or TTS handling that might explain this? 2. Irrelevant / Random Output (“send it”) - The agent is occasionally says unrelated phrases like “send it” without context. Example Call IDs: 019da0ea-6111-7000-8b3b-7158d1a5a7f6 (1:34) 019da0cc-4c5e-7000-adb0-9c47d26122a6 (0:47) 019da0bf-98a8-7000-bd25-ca4b68004c2d (0:10) What could be causing these hallucinated or unintended outputs? Is this related to the LLM, transcription errors, or partial context handling? 3. Overall Call Experience Degradation Latency, response quality, and stability have noticeably worsened. Were there any recent platform updates or changes that could affect agent performance? 4. Configuration Stability - Our configuration was not changed. - Could there be any backend changes affecting previously stable setups? - Are there recommended settings or safeguards to ensure consistent behavior? 5. Recommended Stack (Cost-Constrained) - Could you suggest the best combination of: Transcriber LLM TTS Our target cost is approximately $0.15–$0.18 per minute. We’re aiming for a balance between reliability, latency, and natural conversation quality. This issue is affecting our production workflows, so we’d appreciate a prompt investigation and guidance. We have mailed for a along time and have not received any reply this is extremely unprofessional. I have account with vapi with email: 00prajwal@gmail.com
v
Actionable Next Step: To ensure your issue receives prompt attention, please check the [Vapi Status Page](https://status.vapi.ai/) for any ongoing platform incidents. If no issues are reported, continue the iterative debugging process by testing any changes and sharing new Call IDs with support. This structured approach will help the Vapi team investigate and resolve your production issues efficiently ([Issue Reporting documentation](https://docs.vapi.ai/issue-reporting)). Source: - [Issue Reporting documentation](https://docs.vapi.ai/issue-reporting)
m
I completely understand how frustrating this is, especially when your setup was previously stable. I can help you pinpoint whether the issues stem from TTS streaming, transcription errors, or context handling, and recommend a reliable stack within your cost target. To help me diagnose quickly, can you share if these abrupt stops or random phrases happen across all calls or only specific scenarios? I can work with you directly to get this resolved. @karnatic
k
All calls
It is not following simple names as well
Team can someone please help
This is critical
s
I feel like Vapi’s customer support on Discord has been getting pushed more and more into the background lately. I remember that half a year ago, even my smallest issues were solved within 1–2 hours. Now, it mostly feels like self-proclaimed “gurus” are trying to help and sell their own stuff. I would really love for us developers to have a reliable and fast support system we can count on whenever problems come up. Vapi team, please take this as feedback rather than negativity or bad intent!
s
Hi, Thank you for sharing those call IDs - they made it possible to trace exactly what happened. I've gone through all four calls in detail and have clear answers for each issue you raised.
Issue 1: Agent stopping mid-sentence (call 019da136 at 0:16) Your configuration has not changed and there was no platform outage. What's happening is that Eva's own TTS audio is being picked up by the transcriber, which interprets it as the caller speaking and triggers a barge-in. So when Eva says "that's great! your two-year-old would love our toddler montessori program..." her own voice echoes back into the mic, the system detects audio activity, assumes the caller is interrupting, and cuts her off. The caller then hears silence and says "hello?" - but Eva was never actually interrupted by the caller.
Issue 2: Random phrases like "send it" (call 019da0bf at 0:10) This is the same root cause - echo - not an LLM hallucination or a prompt issue. The transcriber's audio window is still open while Eva is speaking, so it captures both the caller's words and the tail end of Eva's TTS audio bleeding back in. The combined result gets sent to the LLM as if the caller said it, and the LLM responds accordingly. On that specific call, the caller said "just a second" but the transcriber received "just a second, sir. send it." because of the echo.
Issue 3: Overall degradation The call logs show driftDetected events across multiple calls in the 125–150ms range. These contribute to the choppy feeling, but the barge-in false triggers are making it significantly worse because every false interruption resets the pipeline and adds overhead. Fixing the endpointing issue should also improve overall call smoothness.
The fix Increase the endpointing threshold in your transcriber config. Endpointing controls how long the system waits before deciding the user has stopped talking - a lower value makes it more trigger-happy and more likely to false-positive on echo. Try setting it to 300–500: "transcriber": { "provider": "deepgram", "endpointing": 300 } After making that change, test with a few inbound calls by letting Eva finish her full greeting without speaking, then talk naturally. If she stops cutting herself off and the random phrases disappear, the fix is working. If issues persist after that, reply with a new call ID and we'll dig in further. Recommended stack for $0.15–0.18/min
For your use case - inbound phone, natural conversation, cost-sensitive — here is what we'd recommend. For the transcriber, Deepgram Nova-2 or Nova-3 is fast, accurate, handles phone audio well, and supports the endpointing config above. For the LLM, GPT-4o-mini gives you fast response times around 200–400ms and is well suited for conversational flows. For TTS, either Deepgram Aura or ElevenLabs Turbo v2.5 both offer low latency and natural sound within your budget. This stack should land you around $0.12–0.16 per minute depending on call length, likely toward the lower end given how concise your prompts appear to be. I'm sorry for the delay. Please do reply with a new call ID if the endpointing change doesn't fully resolve things and we'll continue from there.
PreviousNext
Powered by