Hi Vapi Team, I’m experiencing a recurring issue on some calls (roughly 3-4 out of every 10) where the assistant does not behave as expected. Here’s what’s happening: The assistant begins the call by playing the “First Message,” but it seems to bypass the LLM entirely — it feels like the assistant isn’t “turned on.” The caller responds, but the assistant either: Takes several seconds to generate the next message, or Says nothing at all, despite the caller continuing to speak. Eventually, the system triggers the “Are you still there?” message and then ends the call after 30 seconds, as if no interaction happened. I’ve checked the call payload, and in these cases there’s no transcript or dialogue history — it looks like no conversation happened at all. It seems like the LLM isn’t being loaded properly on these calls, and the assistant is defaulting to system fallback behavior. Could you please look into this urgently? Call IDs: 5baa70e8-9b5c-4878-bb93-25a06f03265b 33cc60ef-4a86-4c19-92f1-ad83101ba137 8c8129a4-f101-4ea0-a09b-a15a5ceb785f 72926141-3f19-46d7-a55e-1947e0c0d8b7 These 4 happened within 3 minutes. Thanks, Lukasz @User

same here, please have a look on this

Same thing here. We have active clients and this must be fixed asap

@Shubham Bajaj @Sahil

@J. @zurpol please push the button! https://discord.com/channels/1211482211119796234/1293690685781839933

@mmvora @David Jurado @J. @zurpol Looking into it.

thank you 🙏🏼

@zurpol could you request the your telephony team to provide the PCAP file containing RTP packets?

Is there anything I can do right now or should I have my clients turn off their assistants

You can wait for a while until I provide a definitive update.

Just in case it helps, we have disabled the noise filter, and it seems that we haven't received the error in the last few minutes. Not sure if this is the actual cause, but thought it might be useful. We had a similar issue a few months ago as well.

"backgroundDenoisingEnabled": false ?

Ours is off by default

yes, or in our case backgroundSpeechDenoisingPlan: { smartDenoisingPlan: { enabled: false } },

Yours is off and you still have the issue?

now is off

We set it to off a while ago, so we could enable it for individual clients, but we haven't actually started testing it yet

Hi @Shubham Bajaj , @zurpol likely wont have this information. However when I was doing a test call yesterday morning I experienced the exact same issue. I did about 10 calls testing the end-call endpoint and in one of the calls the agent couldn't hear me. Call ID:d59ebe6f-31eb-4206-9ca5-dfc6c77f994d

As @OLEYO said, I don't have this information. Maybe someone who's also experiencing this issue could help?

I can see the call logs in Twilio but I can't see anything to download RTP packets..

me too

@David Jurado @Chris @zurpol @OLEYO @J. @mmvora Can you try few test calls with some other transcriber, from your current choice.

Let me check

update

The problem appears to be coming from the transcriber's side, and we are working with the team to fix it. For now, as a quick solution, could you please use your own key or try a different transcriber.

Can we use a different deepgram model or should we use something else entirely? Thank you for working on this, I appreciate it.

I've tried using different versions of Deepgram (we use Nova 3 version) and different transcribers entirely, but Deepgram is the only one delivering results (booked appointments). The problem doesn't seem to occur with Cartesia, but when using it, the bot isn't booking calls, so it's worthless to the client. Please keep us updated @Shubham Bajaj, or if anyone finds a working solution, please let us know!

Deepgram experienced an unexpected spike in requests, which, caused call failures on your end. To address this, we’ve requested an increase in capacity or the concurrent pool to ensure your calls function smoothly and operations return to normal. After some time you can continue calls without any failure.

Got you, thanks for the update. So, have they increased the capacity/concurrent pool yet, do you know?

You don't have to, we are working with them. In case you are using your own key, then you have to contact the deepgram.

No, I'm not using my own key, so I'll leave it in your capable hands. Could you let us all know once the capacity/concurrent pool has been increased, so we can update our clients? Many thanks in advance!

Hey Zurpol, thanks for sharing your details. Our team will look into it and get back to you soon with an update.