Nvidia has open-sourced PersonaPlex, a voice AI model that listens and speaks at the same time. The 7-billion-parameter model switches speakers in 0.07 seconds, against 1.3 seconds for Google’s Gemini Live in the same tests, and its licence allows commercial use without paying Nvidia a cent.
It never waits its turn
Traditional voice assistants run speech recognition, a language model and speech synthesis in sequence — the polite pause before every answer. PersonaPlex is full-duplex: one model handles both directions at once, updating its state while you speak and streaming its reply immediately. Nvidia says it also learns the behaviour of conversation, not just the words — when to pause, when to interrupt, and when to make the small sounds that signal listening.
The model takes two prompts: a short audio sample that captures a voice, and a text prompt that defines the role. The demos run from a bank teller to a doctor’s receptionist to an astronaut managing a reactor meltdown on a Mars mission. Nvidia’s pitch:
“For the first time, you get both the customization you need and the naturalness that makes conversations feel genuinely human.”
Voice AI, measured
| Measure | PersonaPlex | Gemini Live |
|---|---|---|
| Speaker-switch latency | 0.07s | 1.3s |
| Naturalness score (MOS) | 3.90 | 3.72 |
| Speaker similarity (voice cloning) | 0.57 | ~0 |
The Decoder reports the benchmarks beat Qwen 2.5 Omni and Moshi as well, on naturalness and on speaker-change success (99.2 per cent). The rivals can hold a conversation; PersonaPlex can hold it in a voice you supplied.
Nvidia built the model on Moshi, Kyutai’s open conversational model, then filled the training gap with 7,303 real conversations from the Fisher English corpus and more than 144,000 synthetic assistant and customer-service dialogues. The transcripts came from Qwen3-32B and GPT-OSS-120B, the speech from Resemble AI’s Chatterbox. Training took six hours on eight A100 GPUs — a long lunch by cloud billing standards, not a supercomputer programme.
Free, within limits
The code is MIT-licensed, the weights under Nvidia’s Open Model License: commercial use is allowed and Nvidia claims no rights to outputs. The model is English-only for now, and the researchers’ next steps are post-training alignment and tool integration. Weights and code are on Hugging Face and GitHub, and the model card shows roughly 225,000 downloads in the past month.
It is the latest in Nvidia’s open-model push, alongside the Nemotron 3 Nano, which folded vision and speech into a single small model. The broader thesis — your next computer runs agents, not apps — needs a conversational layer that sounds like a person rather than a walkie-talkie.
What it means for developers in the UAE
The licence is the story for the region. Any UAE startup or enterprise can host its own English-speaking voice agent — customer service, reception, call triage — with no per-seat fees and no customer audio leaving its own infrastructure. The catch is the language: PersonaPlex is English-only, so Arabic deployments will have to wait for a model that speaks the region’s first language. Nvidia’s regional footprint is already building, with its Studio Space Cairo open as a free creative hub for regional creators.
What is PersonaPlex?
PersonaPlex is Nvidia’s open-source, 7-billion-parameter speech model for real-time voice conversations. It listens and speaks at the same time (full-duplex), and it takes a short audio sample plus a text prompt to define its voice and role. Code is MIT-licensed and the weights use Nvidia’s Open Model License.
Is PersonaPlex free to use commercially?
Yes. The code is MIT-licensed and the model weights use Nvidia’s Open Model License, which allows commercial use without Nvidia claiming rights to the outputs. Nvidia describes the model as ready for commercial use.
What hardware do you need to run PersonaPlex?
The model targets Nvidia GPUs — A100 or H100 — and runs on Linux with PyTorch. The BF16 weights come to roughly 16GB of VRAM. Training took six hours on eight A100 GPUs, so a single GPU is enough for running conversations.
Leave a Reply