Overview | NeMo Labs Voice Agent

The large language model (LLM) stage receives the aggregated conversation context and streams the assistant response to text-to-speech. NeMo Voice Agent can run a Hugging Face model in the server process, call a vLLM server, or use a hosted NVIDIA endpoint.

Core Concepts

The following pages explain how language models generate, reason about, and consume voice-agent context.

Concept	What It Explains
LLM backends	Backend selection, configuration layers, and supported tool-calling paths.
Reasoning mode	Thinking configurations and how reasoning text is kept out of speech.
Multimodal and omni models	Direct audio input and the additional user-audio buffer stage.

Model concepts describe runtime behavior. For deployment and endpoint setup, continue to Serving with vLLM, vLLM Plugins, or NVIDIA NIM and Riva.

Use these pages to configure model behavior, serve a backend, or connect models to application tools.

System prompts — control assistant behavior.
Tool calling — let supported backends call functions.