Thanks — I understand you want to know whether LiveKit can make an agent produce short filler sounds like “um / mm-hm / ok” while a customer is speaking, or otherwise give a quick contextual acknowledgement while waiting for an LLM response. Short answer • LiveKit does not have a built-in feature to automatically emit filler interjections during the customer’s speech (agents normally wait until the user turn completes before speaking). • You can, however, implement immediate acknowledgments or non-verbal feedback after the user finishes their turn, and there are ways to simulate responsiveness while the LLM is working. What you can do today • Quick verbal acknowledgments after user turn: ◦ Use the

on_user_turn_completed

lifecycle hook to trigger an immediate short utterance such as

session.say("Okay, let me check that for you.", allow_interruptions=False)

. This runs as soon as the user turn ends and gives the user immediate feedback while longer processing (RAG/tool calls/LLM) continues. • Background / non-verbal feedback during processing: ◦ Use the

BackgroundAudioPlayer

/

thinking_sound

option to play a short “thinking” or typing audio clip while the agent is processing. This provides audible feedback without speaking over the user. • Injecting custom audio (advanced): ◦ If you absolutely need short sounds played while the user is still speaking, it would require a custom implementation using low-latency event hooks and audio injection. LiveKit uses turn detection to determine when the user finished speaking, and agents don’t normally interrupt that flow—so this approach is advanced and not a built-in behavior. Trade-offs to consider • Speaking over the user can degrade turn-detection and natural conversation flow; usually it’s better to wait for end-of-turn and then give a brief acknowledgement. • Short post-turn replies + background “thinking” audio give the perception of responsiveness while avoiding interruptions and are simpler to implement. If you want, I can: • Show a short example of using

on_user_turn_completed

with

session.say()

to produce an immediate acknowledgement. • Describe how to enable and configure `thinking_sound`/`BackgroundAudioPlayer`. • Discuss what “custom audio injection during speech” would require so you can judge feasibility. Which of those would help most for your use case? Sources: Improving voice AI's turn detection with transformers | Pipeline nodes and hooks | LiveKit Documentation | Agent speech and audio | LiveKit Documentation