personaplex

This is a legit finetuning + systems work contribution: full-duplex speech (speak while listening), persona/voice control, and sub-100ms latency aren't trivial. The paper exists (arxiv 2602.06053), weights are published, and the repo has actual inference code—not a wrapper around a closed API. Downsides: finetuned Moshi weights (borrowed credibility), narrow training distribution (synthetic conversations + Fisher corpus), and 'generalization' section is basically "we encourage you to find wei...