NVIDIA · 2026-01-15 · notable
PersonaPlex — NVIDIA's real-time speech-to-speech model with persona control
Real-time full-duplex speech-to-speech model with persona control via text prompts and 16 pre-packaged voice conditionings — built on the Moshi architecture and Helium LM, open-weights under the NVIDIA Open Model license. 9.1k stars.
A real-time speech-in/speech-out model from NVIDIA you can prompt into a specific persona and voice, with CPU offload for low-VRAM machines.
Key specs
| GitHub stars | 9.1k |
|---|---|
| Stars this week | +2.9k |
| Pre packaged voices | 16 |
| Code license | MIT |
| Weights license | NVIDIA Open Model |
What is it?
PersonaPlex is NVIDIA's open release of a full-duplex conversational speech model — it listens and speaks at the same time, handles turn-taking naturally, and can be steered into a specific persona through text prompts or a short voice sample. Code is MIT, weights ship under the NVIDIA Open Model license, and it runs on a single GPU with optional CPU offload.
How does it work?
The model builds on Kyutai's Moshi architecture and integrates the Helium language model for the textual reasoning layer, with a Mimi-style audio codec for the streaming speech path. It was trained on synthetic conversations plus real dialogue from the Fisher English Corpus, giving it 16 pre-packaged voices and prompt slots for assistant, customer-service, and casual-chat roles. Persona drift is controlled by conditioning on both a text prompt and an optional audio reference.
Why does it matter?
Open full-duplex speech models are rare — most voice pipelines are still ASR + LLM + TTS, which costs latency and loses prosody. PersonaPlex is a production-grade, research-quality drop-in for real-time voice agents, with persona control as a first-class feature and open weights you can fine-tune.
Who is it for?
Voice-agent builders, conversational-AI researchers, robotics teams wiring up speech interfaces.