PersonaPlex
Full-duplex speech-to-speech model with prompt-set roles and voices
What it is
A 7B real-time conversational speech model, fine-tuned from Kyutai's Moshi, that listens and talks at the same time and takes a text prompt for its role plus an audio voice embedding for how it sounds. The repo ships inference code, a local web UI server for live talk, and an offline script that turns an input WAV into a spoken reply. Researchers and developers building voice agents can use it to test customer-service or casual-chat personas with interruptions, backchannels, and turn-taking.