Randomly talk to different voice models. Roughly half of the time you should be speaking to Personaplex which is full-duplex, but this can be influenced by giving feedback.