Player not loading? Watch on YouTube
Caroline di Vittorio, a software engineer at LangChain, walks through adding voice to an existing LangGraph AI agent with Pipecat. Her example handles gym membership support through a triage agent and specialists for cancellations, credits and bookings. A FastAPI server tests the text conversation first. She explicitly describes that server as a demo, not production-ready, and uses dummy membership data with real tool calls.
The voice setup replaces the quickstart's OpenAI LLM service with a LangGraph adapter. Pipecat handles speech recognition and speech synthesis. The speaker explains that its interruption handling cuts off the assistant and trims conversation history to what the user heard.
That context handling drives the main architectural change: the voice graph runs without a checkpointer. Instead of persisting the active agent in graph state, it derives the current agent from the latest handoff in message history. The adapter passes Pipecat's messages into the graph and retains tool calls in context without speaking them aloud.
The tutorial then adds a span processor and enables Pipecat tracing for LangSmith. An audio buffer processor supplies recordings, which appear alongside the trace spans. The final advice concerns voice prompts: use shorter replies and ask one question at a time. The demonstrated pipeline does not establish an offline or local LLM deployment.
This is an official LangChain tutorial presented by a member of its engineering team. The voice cancellation run succeeds against dummy membership data, and a second run verifies the expanded tracing.
