Too much backchanneling

Support Help
1

I created my first agent and it’s backchanneling too much. It keeps saying “uh,” “um,” etc.

I’m using Gemini 3 Flash as my AI agent and I also don’t have the option to disable backchanneling under Global Settings > Speech.

I have included this in the global prompt: “Don’t Do Backchanneling: Never use filler words such as ‘um’, ‘like’, ‘uh’, etc.” but this still doesn’t fix it.

I want my agent to sound professional. Kindly help with this.

2

Thank you for reaching out to Retell AI Support. We’ve received your ticket and our team will respond within 8 hours.

3

Hello @admin13,

Can you share the call ID showing this issue?

4

call_5650c1aa00b84ca17741cfc505e and call_214d433bb7128ec618e88414728

There are others too but this ones are latest.

5

Can you try using GPT-4.1 to see if that fixes your issue?
Sometimes these filler words come from the LLM itself

6

I’m having the same issue. I recently moved my architecture from multi-prompt to conversation flow; the multi-prompt agent never did this and the conversation flow agent talks like a 13 year old. Similar prompts. Moving from Gemini 2.5 to 3.0 helps a little but still not deployable. Adding additional prompt instructions doesn’t help at all. I still have my multi-prompt agent live if that helps the retell team.

Please help, Retell.

7

Can you try GPT 4.1?

8

Haven’t tried GPT 4.1 but did try Haiku 4.5 and having the same issue - hoping that you’re not suggesting that I’ll be tied to. GPT 4.1 going forward…FWIW, GPT 4.1 sounds too robotic for my use case so trying to avoid.

9

Can you share the call id for the Haiku 4.5

10

call_ac1fb3c52eab22f2dcc5dc216de

11

Hello @neil,

I just checked this call, and it looks fine.

Didn’t see many filler words like once every few turns, which is normal for a 15-minute call.

13

So there is no way to turn them off? It sounds silly for an AI agent to say “um”. Furthermore, it only started happening when I started to use conversation flow and it’s terrible on Gemini which is my preferred model. Can you help me understand the regression?

14

Hello @neil,

I’ve checked with the team, and this feature is turned off.

These filler words are coming from the LLM itself. What you can do is prompt the LLM to not use these filler words and choose an LLM that will follow the instructions; try using GPT-4.1.

If you mentioned that this model sounds kinda robotic, and if you can share a call ID for this issue, I can see where it is coming from.