Bland’s Tofu-1 is a full-duplex speech-to-speech model; Bland says over 92% of screening listeners couldn’t tell it apart.