Realtime API:"incomplete" response when input is in a non-english language. reason:"content-filter"

API
, ,
1

Hi,
I am trying to generate a translated-transcript of a non-english conversation.
When providing speech-input to OpenAI’s realtime API with a non-english language, the response.done frame is returned as “status:incomplete” in some cases citing “content-filter” as the reason.

{“type”:“incomplete”,“reason”:“content_filter”}

I am sure that the input-audio stream has no content that needs to be stopped by the filter. It is a rather simple straightforward conversation.

-is this happening because of possibly-inaccurate transcription of the non-english language.?
-Is there a way to work around this content-filter and a non-english language.? How accurate is the content-filter in itself with a non-english language.?
-am i the only one facing this?

2

I am also using the Realtime API dealing with {“type”:“incomplete”,“reason”:“content_filter”} on response.done. However, my incomplete status seems to happen only when I add instructions that are more than a few sentences long (well under 100 tokens). I am changing the instructions on response.create.

@mrenga.96, are you adding custom instructions to the response (or maybe the session)?

I don’t understand why my custom instructions are creating this error (the instructions are very generic and should not be causing any issues with the content_filter).

3

Yes. I simply mention the language that the input is in. I havent had any luck working around this. Have you.?

4

Same problem here.
Found no solution yet, though.