Overview
- Block revealed Dessa on October 23, 2026, posting demos and technical notes that show real‑time, back‑and‑forth audio conversations between the model and people.
- The team says Dessa was trained only on human audio with no text data, a choice meant to keep timing, prosody and emotional cues that text‑based pipelines remove.
- Dessa runs in full‑duplex, which lets it listen while it speaks so it can interrupt, acknowledge, laugh and manage conversational timing like a phone call.
- Block says the model can run locally on a MacBook Pro M4 Max or stronger hardware and will be given first to a small group of AI safety researchers for study.
- Public demos look mostly natural but include awkward errors such as misnaming a speaker, prompting concerns that convincingly human voices could enable impersonation, make verification harder, and change how people trust voice agents.