The equivalent to

session.generate_reply

in realtime models is still

session.generate_reply()

. This method works with both traditional LLM pipelines and realtime models, allowing you to prompt the model to generate a response by providing

instructions

and/or

user_input

. The response is returned as a

SpeechHandle

object, which lets you track speech state and handle events such as speech completion or interruption. For scripted speech (the equivalent of

session.say()

), realtime models don't natively support direct TTS output unless you add a TTS plugin to your session. If you need to play back a specific text, you must add a TTS plugin and use `session.say()`—otherwise, use

generate_reply()

and provide explicit instructions, though the model output may vary. • For precise scripting with realtime models, use a TTS plugin with

session.say()

. • For dynamic replies or LLM-driven generation, use

session.generate_reply()

. Both methods return a

SpeechHandle

for tracking and synchronization. For further details, refer to your platform's plugin and session management documentation. If you need practical implementation examples for either approach, let me know your use case. Sources: Agent speech and audio | LiveKit Docs | tests/test_agent_session.py | Realtime model integrations | LiveKit Docs