Block's Dessa is a full-duplex speech model trained only on speech, no text. What Block shows, what it has not released and how it compares.