Block’s Dessa Promises Voice AI That Can Hear More Than Words

Block’s Dessa Promises Voice AI That Can Hear More Than Words
Block’s latest AI experiment began with a deliberately simple premise: teach a model to speak the way people do, rather than funneling every conversation through text.

The company, founded by Jack Dorsey, has released details of Dessa, a full-duplex voice model that can listen and respond at the same time. Block says it trained the system entirely on human speech and designed it to carry laughter, pauses, fatigue and distress as part of the voice itself—not as instructions bolted onto a text pipeline. “The most powerful AIs in the world can’t hear how you feel,” Block argued, calling that limitation both a poorer experience and “potentially dangerous.”

That argument is also Block’s pitch for why Dessa matters. The company says many current voice systems discard information when they convert speech to text, while its model learns from conversational audio. It says Dessa can run locally on a consumer machine, including a MacBook M4 Max or better, potentially lowering the cost and latency of voice products. For a company already using voice AI in Square’s ordering tools, the commercial implications are clear.

But the early public material also shows the tension in making an assistant sound more like a person. Block’s demonstrations feature varied voices and cadences, including a model that sounds tired after discussing work and another that appears close to tears. The Verge described the exchanges as “mostly natural” but noted that Dessa at one point mistakenly called its conversation partner “Ronaldo,” capturing the oddness that can break through an otherwise fluid chat.

For now, Dessa is not publicly available. Block has instead invited safety researchers to probe the model—a significant caveat for technology that may be persuasive enough to confuse people on the other end of a call. The company presents that scrutiny as part of its goal; the demos suggest it will be central to whether emotionally fluent voice AI earns trust.

https://foxvector.com