Pipeline Nodes, Hooks & Custom Logic

Control every stage of the voice pipeline

Override every stage of the voice pipeline: STT preprocessing, structured LLM output, emotional TTS, RAG injection, content filtering, and realtime model nodes.

What You Build

Chain-of-thought agent with structured JSON output, emotional TTS, content filtering, RAG injection, and full pipeline customization.

Prerequisites

→Course 1.1
→Course 2.1

Chapters

The voice pipeline architecture

20m

Full pipeline diagram, async generators, and how Agent.default wires everything together.

Pipeline diagramAsync generatorsAgent.default

stt_node: audio preprocessing

20m

Override stt_node to preprocess audio, filter noise, or detect keywords before transcription.

stt_nodeAudio filteringKeyword detection

llm_node: structured output

25m

Override llm_node for structured JSON output with TypedDict schemas and streaming parsing.

llm_nodeResponseEmotionStreaming JSONresponse_format

tts_node: emotional speech

20m

Override tts_node for emotional speech with dynamic TTS instructions and pronunciation control.

tts_nodeTTS instructionsPronunciationVolume

on_user_turn_completed: RAG injection

25m

Inject RAG context into the conversation by overriding on_user_turn_completed.

on_user_turn_completedturn_ctxStopResponseRAG

Content filtering

20m

Build content filters: simple keyword-based and LLM-powered, plus transcription node overrides.

Content filterLLM filtertranscription_node

Realtime model nodes

20m

Override realtime model nodes for custom audio output processing.

realtime_audio_output_nodeRealtime models

Full assembly

20m

Assemble the complete chain-of-thought agent with all customizations, tests, and metrics.

Complete pipelineTestsMetrics

What You Walk Away With

Complete mastery of LiveKit's voice pipeline architecture. Ability to customize any stage for specialized use cases.