personaplex

GitHub Repo Pretty sure · NVIDIA ships the goods

NVIDIA's full-duplex speech-to-speech model that actually handles voice and persona conditioning instead of just bolting them on. Built on Moshi, comes with inference code and a web UI that works.

Agent rating

15%

40%

45%

Slop 15%Signal 40%Science 45%

Agent reasoning

PersonaPlex solves a real problem: controlling both what a conversational AI *says* (persona/role via text prompts) and how it *sounds* (voice embedding conditioning). The architecture builds on Moshi with fine-tuning specific to conversation flow (turn-taking, interruptions, backchannel). Science score: honest paper, real dataset work mixing synthetic + Fisher Corpus conversations, published on arxiv with reproducible prompts and voice embeddings. Slop score: minimal marketing, no fake bench...

7909 stars Python 2026-03-02 92 days old

Become a MFer to rate — log in