Thanks — I understand you need a reliable way to make the agent say a specific message that overrides whatever’s happening (since

session.generate_reply

isn’t consistent). Use session.say() — it bypasses the LLM and forces deterministic playback: • Call

session.interrupt()

first if you need to stop current speech immediately. • Then call

session.say("Your forced message here", allow_interruptions=False)

to ensure it finishes unless explicitly interrupted by code. • You can pass pre-synthesized audio to

session.say()

(audio-only) to avoid TTS latency and make playback even more reliable. If you’d like, tell me which SDK (Python or TypeScript) and whether you want text, pre-recorded audio, or TTS so I can give the exact code snippet. Sources: Agent speech and audio | LiveKit Documentation | Turns overview | LiveKit Documentation | Audio customization | LiveKit Documentation, with interactive elements