NVIDIA PersonaPlex-7B: Open Source Full-Duplex Voice AI

What Happened

NVIDIA has released PersonaPlex-7B-v1, a 7 billion parameter speech-to-speech model that fundamentally changes how voice AI handles conversation. Unlike every voice assistant you have used before, PersonaPlex does not wait for you to finish talking before it starts responding. It listens and speaks at the same time.

This is called full-duplex interaction, and it is the same way humans naturally converse. You can interrupt it mid-sentence, and it adapts. It produces backchannels like “uh-huh” and “oh, okay” while you are still speaking. It pauses when appropriate. No rigid turn-taking. No awkward silence while the AI processes your words.

Why Traditional Voice AI Falls Short

Traditional voice assistants run a three-stage pipeline that creates an unnatural conversation flow:

Each stage adds delay, and the system cannot hear you while it is generating a response. This is why conversations with Siri, Alexa, or Google Assistant feel robotic. You speak, wait, get a response, speak again.

PersonaPlex replaces this entire pipeline with a single Transformer model that processes incoming audio and generates speech simultaneously.

Core Capabilities

How PersonaPlex Works

Dual-Stream Architecture

PersonaPlex is built on the Moshi architecture from Kyutai, with Helium as the underlying language model backbone. The architecture uses two parallel streams:

• User stream - continuously encodes incoming audio from the user’s microphone
• Agent stream - simultaneously generates the AI’s speech and text response

Both streams share the same model state. This means PersonaPlex can adjust its response in real time as the user speaks, enabling barge-in, overlapping speech, rapid turn-taking, and contextual backchannels.

The Mimi neural audio codec handles audio encoding and decoding at 24 kHz, converting waveforms into discrete tokens that the Transformer can process.

Hybrid Persona Control

PersonaPlex uses two inputs to define conversational identity:

• Text prompt - describes the role, background, organization, and conversation context (up to 200 tokens)
• Voice prompt - an audio embedding that captures vocal characteristics, speaking style, accent, and prosody

This hybrid approach lets you create a customer service agent for a specific company with a specific voice, a wise teacher who sounds warm and patient, or a fantasy character with dramatic inflection. The persona stays consistent throughout the entire conversation.

Demonstrated Personas

Benchmark Results

NVIDIA evaluated PersonaPlex on FullDuplexBench and a new extension called ServiceDuplexBench for customer service scenarios. The results show clear advantages over both open-source and commercial alternatives.

Conversational Dynamics

Latency

Task Adherence

PersonaPlex is the only model that scores above 4.0 on both benchmarks, combining strong general knowledge with reliable task-following in structured business scenarios.

Training: Less Than 5,000 Hours

PersonaPlex was trained in a single stage using a carefully designed blend of real and synthetic conversations.

Real Conversations

7,303 calls (1,217 hours) from the Fisher English corpus provided natural conversational patterns - backchannels, disfluencies, emotional responses, and authentic turn-taking behavior. These recordings were back-annotated with persona prompts using GPT-OSS-120B at varying levels of detail.

Synthetic Conversations

• 39,322 assistant dialogs (410 hours) - generated with Qwen3-32B and GPT-OSS-120B, synthesized to audio with Chatterbox TTS from Resemble AI
• 105,410 customer service dialogs (1,840 hours) - covering various business scenarios with structured prompts including company names, pricing, and operational rules

The training design disentangles two qualities: naturalness from real conversations and task adherence from synthetic scenarios. The hybrid prompt format bridges both data sources, letting the model combine natural speech patterns with precise instruction following.

What This Means for Voice AI

PersonaPlex represents a significant shift in what open-source voice AI can do. Until now, the choice was between customizable but robotic cascaded systems and natural but inflexible full-duplex models. PersonaPlex eliminates that trade-off.

For Developers

The model is ready for commercial use. Developers building voice agents, customer service bots, or interactive characters now have an open-source foundation that rivals proprietary systems. The MIT-licensed code means full freedom to modify and deploy.

For the Voice AI Industry

Full-duplex interaction has been the holy grail of conversational AI. Google, OpenAI, and others have invested heavily in making voice assistants feel more natural. NVIDIA has now open-sourced a model that achieves this at the 7B parameter scale, lowering the barrier for anyone to build truly conversational voice interfaces.

For Creators and Businesses

Voice-first interfaces are accelerating across customer service, accessibility tools, gaming, and content creation. PersonaPlex’s persona control makes it practical for specific business use cases where the AI needs to sound on-brand and follow structured scripts while still feeling human.

Current Limitations

• English only - no multilingual support yet
• Requires NVIDIA GPUs - optimized for Ampere and Hopper architectures (A100, H100)
• Limited training data - under 5,000 hours, which may restrict performance in niche dialects or specialized domains
• No production safety testing - NVIDIA notes that bias, explainability, and privacy concerns need additional testing before production deployment

How to Get Started

Quick Start (5 minutes)

Requires a Linux machine with an NVIDIA GPU (Ampere or Hopper) and Python installed.

1. Install the audio codec and clone the repo:

# Ubuntu/Debian
sudo apt install libopus-dev

# Clone and install
git clone https://github.com/NVIDIA/personaplex.git
cd personaplex
pip install moshi/.

2. Accept the model license on Hugging Face, then set your token:

export HF_TOKEN=your_token_here

3. Launch the server (auto-generates temporary SSL certs):

SSL_DIR=$(mktemp -d); python -m moshi.server --ssl "$SSL_DIR"

4. Open https://localhost:8998 in your browser. Start talking — PersonaPlex responds in real time.

FAQ

---

Sources

1. NVIDIA ADLR - PersonaPlex: Natural Conversational AI With Any Role and Voice
2. MarkTechPost - NVIDIA Releases PersonaPlex-7B-v1
3. NVIDIA PersonaPlex-7B-v1 on Hugging Face
4. PersonaPlex GitHub Repository