NVIDIA publishes PersonaPlex speech-to-speech model

PersonaPlex can generate speech while incoming audio is still being heard, allowing interruptions, overlaps and rapid turn-taking.[3] Users can condition the model with both a voice prompt and a text prompt; the voice prompt sets vocal characteristics and speaking style.[3] NVIDIA says PersonaPlex is based on Kyutai’s Moshi architecture and has 7 billion parameters.[1] Its underlying Helium language model is used for semantic understanding and generalization to scenarios outside its training distribution, according to NVIDIA.[1]

NVIDIA published its research announcement on January 15, 2026, while PersonaPlex weights appeared on Hugging Face and code on GitHub the same day.[1][2] The code is released under the MIT License, while the model weights are under the NVIDIA Open Model License.[4] NVIDIA describes the model as ready for commercial use.[3] However, accessing the model files on Hugging Face requires users to agree to share their contact information.[3] The model card lists English as the use case for both input and output speech, and specifies a 24 kHz audio sample rate.[3]

NVIDIA reports training PersonaPlex on 39,322 synthetic assistant-role conversations totaling 410 hours, plus 105,410 synthetic customer-service conversations totaling 1,840 hours.[1] The transcripts were generated with Qwen3-32B and GPT-OSS-120B, and the speech with Chatterbox TTS; NVIDIA says training took six hours on eight A100 GPUs.[1][5] The model card quotes smooth-turn-taking and user-interruption metrics of 0.908 and 0.950, with respective latencies of 0.170 and 0.240 seconds.[3] NVIDIA says PersonaPlex outperforms other open-source and commercial systems on conversational dynamics, latency and task adherence, while its research lists Gemini Live as an exception to its comparison claim.[3][5]

NVIDIA’s research lists scores of 3.90 for PersonaPlex, 3.72 for Gemini, 3.70 for Qwen-2.5-Omni and 3.11 for Moshi.[5] The research also describes Service-Duplex-Bench as 50 service-role scenarios, each with seven questions.[5] These figures provide comparison points, but the evidence does not establish how the reported results translate to a particular organization’s live conversations. PersonaPlex can continue generating speech while listening to incoming audio, and its model card lists English speech for both input and output.[3] NVIDIA says future work will explore post-training alignment and integration with external tools.[5]