v1.0.0

Nemotron Voice Agent v1.0.0 (3 March 2026)

Initial release of Nemotron Voice Agent — an end-to-end voice agent blueprint powered by NVIDIA Nemotron ASR, LLM, and TTS, designed for scalable, production-ready deployments.

Added

• End-to-end voice agent pipeline with NVIDIA Nemotron ASR, LLM, and TTS, supporting streaming audio and mid-conversation interruptions
• Built on the open source Pipecat-ai and nvidia-pipecat frameworks
• NVIDIA Nemotron Speech models:
  • Parakeet CTC 1.1B (English ASR)
  • Parakeet 1.1B RNNT (Multilingual ASR)
  • Magpie TTS Multilingual
• NVIDIA Nemotron LLMs via NVIDIA NIM:
  • Nemotron 3 Nano 30B A3B
  • Llama 3.3 Nemotron Super 49B v1.5
• WebRTC transport for real-time, low-latency voice communication with a custom frontend UI
• Docker Compose deployment with optional TURN server support for remote access
• Multilingual support with automatic language detection and seamless mid-conversation language switching
• Jetson Thor edge deployment support
• Pipeline customizations using environment variables and config files
  • ASR, LLM, TTS model change
  • Speculative speech processing enable/disable
  • Conversation history thresholds
  • output audio buffering
• Open telemetry tracing and monitoring support
• Documentation:
  • Getting started guide covering prerequisites, GPU configuration, and step-by-step setup
  • Configuration guide for pipeline customizations
  • Jetson Thor deployment guide for edge use cases
  • Best practices guide covering production deployment, latency optimization, and conversational UX
• AI agent deployment skill for Cursor and Claude Code to streamline deployment on workstations and Jetson Thor

Known Issues

• ASR transcription can occasionally be inaccurate, though the LLM generally compensates by inferring meaning from context.
• The context aggregator limits chat history to 20 turns by default. Older turns are dropped when this limit is reached, rather than summarized.