gpt-realtime-1.5 leaks audio control tokens (<|audio_text|>, <|caption_quality_N|>) into text stream when run with modalities=["text"]

Welcome to the dev community, @zaidkaraymeh and thanks for the detailed repro, that’s really helpful. If you’re able to share a request ID (and approximate timestamp) for one of these calls, it would make it much easier for the team to trace what’s happening internally. A short raw snippet of the streamed output (like the one you included) is perfect as well. ~Smith