AI · AI model
PersonaPlex-7B (nvidia/personaplex-7b-v1)
Open full-duplex speech-to-speech conversational AI model by NVIDIA
PersonaPlex-7B is a 7-billion-parameter speech-to-speech conversational model developed by NVIDIA that supports full-duplex interaction. It processes continuous audio streams using a dual-stream architecture, allowing it to listen and speak at the same time without the latency of separate ASR, LLM, and TTS pipelines.
What it's for
- Build low-latency voice assistants that support interruptions and natural turn-taking
- Add expressive, real-time voice interaction with emotional nuance to applications
- Study behavioral edge cases and failure modes in speech-to-speech conversational models
Features
- Full-duplex audio processing enabling simultaneous speaking and listening
- Single-model architecture replacing traditional ASR, LLM, and TTS chains
- Support for conversational interruptions, laughter, and backchanneling
- Compatible with the Moshi speech-text framework
- Low-latency streaming audio-to-audio generation
Pricing
Open source. Free download under a custom NVIDIA license
Details
- Type
- AI model
- Category
- AI
- Topic
- Speech Models
- Open source
- Yes
- Platforms
- Linux, Self-hosted, Code library
- For
- AI researchers, speech engineers, and voice agent developers
- Alternative to
- OpenAI Advanced Voice Mode, Kyutai Moshi
- Tags
- speech to speech, full duplex, voice ai, nvidia, audio, open weights
Why it was saved
- @aidailyadvances: Demonstrates full-duplex voice synthesis and the potential failure modes or uncanny behavior of conversational voice models when caught in emotional or repetitive generation loops.
- @ai_with_saqi: Eliminates the rigid, high-latency ASR→LLM→TTS pipeline, enabling human-like interruptions, laughter, and simultaneous backchanneling in voice AI applications.
- @aisearchio: Open-source speech-to-speech model capable of real-time dialogue including jokes and banter, shown in a live two-party voice demo.
Related
Ultra-low-latency streaming text-to-speech and instant voice cloning model.
Open-source real-time multilingual speech-to-speech translation model with voice preservation
Open-weights streaming text-to-speech model with ultra-low latency for English and French
NVIDIA alignment fine-tune of Llama 3.1 70B for enhanced instruction following
Open-source 600M English speech-to-text model with fast inference and timestamps
AI voice generator, voice cloning, and conversational agent platform
Updated