Have you noticed that, with AI With voice assistants, do you always have to “take turns speaking”? You finish a sentence, wait for it to respond, and only after it finishes can you keep going. This “walkie-talkie mode” feels worlds apart from talking to a real person. After all, in everyday human conversations, people constantly interrupt, chime in, or talk over each other. And… NVIDIA Released in January 2026open sourcemodel PersonaPlex-7BThis is precisely to completely break through this restriction.
NVIDIA Releases PersonaPlex-7B: A 100% Open-Source Full-Duplex Voice AI Model
PersonaPlex-7B is a 7-billion-parameter model.Full-Duplex Voice AI ModelBased on the Moshi architecture developed by the French Kyutai team, it uses a dual-stream Transformer design. Traditional voice AI operates as a three-stage pipeline: first, automatic speech recognition (ASR) converts your speech into text; then a large language model (LLM) generates a reply in text; finally, text-to-speech (TTS) reads it aloud. These three steps combined often add up to more than one second of latency, not to mention the information loss between each module.
PersonaPlex-7B takes these three stagesMerge into a single end-to-end model—It takes audio in directly and outputs audio directly, without any text conversion in between. This isn’t just an engineering simplification; it’s a revolution at the architecture level. And it is… 100% open sourceDone. The model weights have been released. HuggingFaceAnyone can download and use it for free. In the first month after its release, downloads exceeded 330,000.
Core Technology Breakthrough: True Full-Duplex Dialogue
The core technological breakthrough of PersonaPlex-7B is…Full-duplex communicationThe model can listen and speak at the same time, without waiting for the other person to finish before responding.
NVIDIA just dropped PersonaPlex-7B 🤯
A full-duplex voice model that listens and talks at the same time.
No pauses. No turn-taking. Real conversation.100% open source. Free.
Voice AI just leveled up.https://t.co/YfzFQfBzMS pic.twitter.com/bVwJ5EFJFB— Hugging Models (@HuggingModels) February 15, 2026
This sounds simple, but is extremely difficult to implement. The model must continuously listen to the user’s audio input while it is “speaking,” determining whether the other party is interrupting, agreeing, or if it is just background noise. PersonaPlex performs surprisingly well in this regard:
- Conversation switching latency as low as 0.170 seconds:You can hardly feel any lag.
- Interrupt processing delay: 0.240 seconds:The moment you cut in, it immediately stops to listen.
- In the FullDuplexBench benchmark test,Interrupt success rate reached 100%
- Support natural backchannel words, such as “uh-huh,” “right,” and “I see.”
In human evaluations of conversational naturalness, PersonaPlex achieved 3.90 points(Out of 5 points), it surpasses Google Gemini Live’s score of 3.72. In other words, chatting with PersonaPlex feels more natural than Google’s current best voice AI.
Hybrid Notification System: Both Sound and Character Are Customizable
Another major highlight of PersonaPlex is itsHybrid Prompting Systemcombining two control methods:
- Voice PromptWith just an audio sample, the model can learn and mimic the voice’s timbre, intonation, and speaking style.
- Text PromptDescribe the persona setting in natural language, for example, “You are a patient banking customer service representative” or “You are an astronaut at a Mars base communicating with Earth.”
This means developers can create voice AI characters for a wide range of scenarios—bank customer service, medical receptionists, online teachers, and even NPCs in games—while maintaining a natural, full-duplex conversational experience. This is an industry first.First to deliver both realistic dialogue and character control.open source solution.
Training data: a hybrid strategy combining real human conversations and synthetic data
PersonaPlex’s training data combines two major sources:
- 7,303 real human conversations, totaling 1,217 hours, from the classic Fisher English corpus. These data provide authentic speech patterns—including hesitations, overlaps, interruptions, and other natural conversation characteristics.
- Over 140,000 synthetic dialogues, covering task scenarios such as customer service and assistant roles. These data enhanced the model’s instruction-following ability on specific tasks.
This “learning naturalness from real-human data and learning tasks from synthetic data” hybrid strategy allows the model to both converse like a human and accurately complete assigned tasks.
Hardware requirements and cost: A single A100 is enough.
One major advantage of PersonaPlex-7B isThe hardware requirements are relatively modest.A single NVIDIA A100 GPU can run the full model, with self-hosting costs ranging from approximately $0.50 to $2.00 per hour.
In contrast, OpenAI’s Realtime API is priced at $0.06 per minute for input and $0.24 per minute for output. For customer service centers with high call volumes, a self-hosted PersonaPlex setup offers a clear cost advantage.
Comparison with competitors: each has its strengths.
Currently, the major players in the voice AI market each have their own strengths:
- vs ChatGPT VoiceOpenAI’s voice mode is still half-duplex (turn-based), so it can’t truly listen and speak at the same time. PersonaPlex has a structural advantage in conversational naturalness.
- vs Gemini LiveGoogle’s solution partially supports full duplex, butNot open source.developers cannot deploy or modify it themselves
- vs ElevenLabsElevenLabs offers over 10,000 voice options and boasts industry-leading audio quality, but its underlying architecture is still round-robin.
- vs MoshiKyutai’s Moshi is a pioneer in full-duplex architecture, and PersonaPlex is built on that architecture, but Moshi lacks character customization and voice cloning features.
Current limitations
Despite the exciting technical breakthroughs, PersonaPlex-7B still has several notable limitations:
- The sound quality leans toward a telephone-like quality.A 24kHz sampling rate feels insufficient by 2026, still falling short of the audio quality offered by commercial solutions like ElevenLabs.
- Limited language supportEnglish currently delivers the strongest performance; other languages have not yet been fully optimized.
- Need GPU infrastructure.No official cloud hosting service; NVIDIA does not provide a hosted API.
- Still in the research stageNot yet fully suitable for direct deployment into a production environment.
Industry Impact: NVIDIA’s “Free Models, Sell GPUs” Strategy
Behind PersonaPlex-7B’s open-source strategy is NVIDIA’s familiar ecosystem play. NVIDIA doesn’t make money from the model itself; it makes money fromThe GPU hardware required to run these modelsAs more and more startups and enterprises adopt PersonaPlex to build their own voice AI services, every one of them needs to purchase or rent NVIDIA GPUs. This is a carefully designed flywheel effect:
- Free open-source models → lower the barrier to voice AI development
- More developers adopt → voice AI APIs become commodities.
- Profits for commercial voice APIs are being squeezed → value is migrating from the application layer to the hardware layer.
- Every self-hosted team is a potential NVIDIA GPU customer.
This follows the same logic as Meta’s open-sourcing of the LLaMA series:When models are free, the shovel sellers are the biggest winners.。
Perspective
The greatest significance of PersonaPlex-7B may lie not in the “full-duplex” technological breakthrough itself, but in the industry trend it represents:Voice AI is being commoditized.In the past, building a natural voice conversation system required stitching together ASR, LLM, and TTS APIs from three different providers; latency tuning and error handling alone were enough to deter small teams. Now, an open-source end-to-end model plus a single GPU can get the job done. The impact on the entire voice AI ecosystem is profound.
However, its limitations must also be acknowledged. 24kHz audio quality is still insufficient for commercial scenarios that demand a brand image; support for languages other than English remains immature; and the lack of cloud hosting means teams without GPU resources are still left out.In the near term, PersonaPlex is better suited as a research foundation and proof of concept than as a direct replacement for existing commercial voice solutions.
For Taiwanese developers, the most critical question is probably:When will Chinese support arrive?Based on the current training data, the model relies almost entirely on English-language corpora; the Chinese full-duplex conversational experience still needs further fine-tuning by the community or local teams. But that is the nature of open source: the code and weights are there, and anyone can take over. Either way, when NVIDIA turns full-duplex voice AI into something you can “download for free,” the rules of the game in this space have changed.
Source: KOCPC Chinese