You can use

session.say()

and

session.generate_reply()

directly inside a function tool. Here are the key patterns and considerations: • Direct Calls: Within your

@function_tool

-decorated function, use

await self.session.say("Your message here")

for predefined messages, or

self.session.generate_reply(instructions="...")

for dynamic, LLM-generated responses. • Automatic tool_choice for generate_reply: When you call

generate_reply()

inside a function tool, the

tool_choice

is automatically set to

none

. This prevents further nested or cascading tool invocations during the reply, ensuring safe usage. • Session Access: Use

self.session

to access speech methods inside the function tool implementation. • TTS Requirement:

session.say()

requires an available TTS plugin. For realtime models, attach a TTS plugin or use

generate_reply()

if unavailable. • Speech Handles: Both methods return

SpeechHandle

objects which you can use to control or await playback as needed. • Error Handling: Calling these methods when the session is closing or not running raises a

RuntimeError

. • Interruption: If using a RealtimeModel with server-side turn detection,

allow_interruptions

cannot be

False

with

VoiceAgent.say()

. This will be adjusted automatically. Example pattern:

@function_tool
async def my_function(self, context: RunContext):
    await self.session.say("Processing your request...")
    self.session.generate_reply(instructions="Instruct the user...")
    return "Done"

These methods enable rich voice feedback and dynamic dialogue within your agent's tools. Sources: Agent speech and audio | LiveKit Docs | tests/test_agent_session.py