Crazy Hallucinations by Realtime API unrecorded in transcript

Has anyone noticed the API over the last week or so doing crazy hallucinations that also don’t get recorded in the transcript at all. I havent managed to setup a second transcription service to capture it yet for full display, but I’ve noticed completely random diatribes about nothing uncaptured. Most recently (5 mins ago) I tried to fake order bratwurst “can i get 2 bratwurst” to test function calling in a phone ordering system and the model started to tell me about the race relations in the US…

I am not from the US or trying to make some political stance this is really what it was talking about out of nowhere

This seems to only happen with response.create… heres the transcript clip:

Conrad York: Um, can I get two bratwurst sausages

(really long completely undocumented 30 second diatribe on race in the USA)

Conrad York: Uh, bratwurst sausages.

AI Agent: Certainly! How many units of Bratwurst sausages would you like to order? They’re available in 500-gram portions.

1 Like

Been seeing this for a couple of weeks and it’s really concerning how irrelevant and random and blatantly incorrect statements it continues with even after the client side receives the response.done event.

It seems that normally, the undocumented “output_audio_buffer.audio_stopped” comes very soon after “response.done” but in the cases where it starts hallucinating, “output_audio_buffer.audio_stopped” comes a long time after “response.done” and like you saw, response.done’s content doesnt contain the hallucinated audio transcription.

I am honestly so confused, it will talk about irrelevant things, and I have even heard what I think is some weird music or background noise going on

I second this. Terrible. In my case it’s “thinking out loud what it has to do”

Just adding that I’ve seen this as well.