Overview
Run NVIDIA Nemotron 3 VoiceChat on Crusoe. An end-to-end speech-to-speech model for real-time voice agents — no separate transcription or synthesis pipeline.
Explore additional models
Frequently
asked questions
Nemotron 3 VoiceChat is NVIDIA’s end-to-end speech-to-speech model, which unifies speech recognition, language modeling, and speech synthesis in a single model rather than a three-model pipeline. It takes spoken audio in and returns spoken audio out, supports a 131,072-token context, and is built on a 12B Nemotron Nano language backbone. On Crusoe it is available for high-performance Serverless Inference.
Nemotron 3 VoiceChat is the only Nemotron model on Crusoe that returns spoken audio. The others answer in text: Nemotron 3.5 Lightning 30B A3B handles text-only workloads, Nemotron 3 Nano Omni and NVIDIA Nemotron 3.5 Super VL also accept image input, and Nemotron 3.5 Parse runs on a dedicated endpoint with a 16K context. Choose VoiceChat for live voice interfaces, and the others for analysis or text-based agents.
Nemotron 3 VoiceChat is served through Serverless Inference, but access and pricing are arranged with the Crusoe team rather than self-serve per-token billing, because voice workloads need sustained real-time capacity. Talk with our team to scope access for your expected call volume and latency requirements.
The largest latency reduction in a voice agent comes from replacing the ASR-to-LLM-to-TTS pipeline with a single speech-to-speech model, which removes two handoffs and the buffering between them while preserving prosody and turn-taking cues. Nemotron 3 VoiceChat is built for that pattern, with a sub-300ms response target.
Nemotron 3 VoiceChat is available for Serverless Inference, with access arranged through the Crusoe team. It is not currently offered for Self-Serve Deployments or Serverless Fine-Tuning. Talk with our team to scope a dedicated, SLA-backed endpoint through Tailored Deployments.
Nemotron 3 VoiceChat is good for real-time voice agents: customer support lines, in-app voice assistants, drive-through and kiosk ordering, and interactive applications that need a spoken answer in the moment. Its 131,072-token context keeps long conversations coherent, and its compact 12B size suits sustained real-time serving.


%204.png)
