Thanks — I’m sorry you’re hitting that. You said you set GPT-Realtime-1.5 to text-only but it still responds with audio; here’s how we can fix or narrow this down. First, confirm how you set modalities (reply with the snippet if you can). Common fixes: • Ensure the model is instantiated with

modalities=["text"]

(not the default

[ "text", "audio" ]

). • Explicitly set or disable TTS in the session. If the session includes a

tts

plugin or parameter, the model may still produce audio. To prevent the model from generating audio, remove the

tts

parameter or set it so audio is handled by a separate TTS service instead of the realtime model. • If you’re loading conversation history, try a fresh session with

modalities=["text"]

(history can sometimes cause inconsistent outputs). Example pattern to keep the model text-only:

session = AgentSession(
    llm=openai.realtime.RealtimeModel(modalities=["text"]),
    # no tts parameter here — or set tts to a separate TTS service only if you want audio
)

If you can paste the exact code/JSON you’re using to create the realtime session (the

RealtimeModel

or

AgentSession

snippet), I’ll point out what to change. Sources: Realtime models overview | LiveKit Documentation, with interactive elements