No audio output after manually adding conversation items

API
, ,
1

I’ve noticed that if I manually create conversation items with the realtime api, the api will stop generating audio responses and only respond w/text (despite having modalities set to text and audio).

I also noticed that pipecat identified the same issue.

I’m curious if others have observed this? Is there a workaround other than placing all of the context into the first conversation item?

1 Like
2

Hey!

Just putting my answer from another post here:

Please check out the related posts below this very post.

Cheers! :hugs:

2 Likes
3

Thanks. Including both modalities doesn’t seem to help if I manually add a bunch of text conversation items. That post is very relevant though to my problem.

2 Likes
4

Hey @dlite, what did you end up doing?

I am using the “gpt-4o-audio” model and seems to have the same issue. After adding conversation with long text, the model doesn’t seem to output any audio object…

5

I think a work-around is adding the conversation history to the system message (vs. appending actual conversation items).

6

Thanks mate, I end up using two stages approach: audio-preview → ttl model (text to speech) given my sufficient funding.