All stories
Block Dessa Launches Developing

Block's Dessa skips text in its pitch for more natural voice AI

The company describes a speech-trained, full-duplex model, with demonstrations available before broad public access.

Sources (2)

Why now October 23 announcement, reported October 24. Backfilled October 27 after a publication gap; this site's publication time is not the event date.

Block is pitching Dessa as a different route to conversational voice AI: a model trained from scratch on human conversation rather than built around a text model. Its October 23 announcement, covered by TechCrunch the following day, emphasizes timing, emotion, pauses and the ability to listen while speaking.

Block describes the system as full duplex. That makes its demonstration relevant to people frustrated by voice assistants that force a rigid speak-then-wait exchange. The company’s page presents sample conversations as live interactions; those are company demonstrations, not sessions independently reproduced by AITrending.

TechCrunch reports that Block says the model can run on a consumer MacBook with an M4 Max chip or better. That is an attributed hardware claim, not a measured performance result from this publication. Speech-only training also does not establish human understanding, consciousness or equivalent factual reliability.

Access is an important boundary. The reporting describes limited availability and an invitation for safety researchers, rather than a generally downloadable model. Readers should treat Dessa as a promising voice demonstration with a distinct training approach, not an available replacement they can already install in any assistant. Public weights and broad deployment were not established by the reviewed sources.