Block has unveiled Dessa, a speech-to-speech AI model that it says was trained from scratch on human conversation and designed to avoid a familiar flaw in voice assistants: reducing a person’s voice to plain text before responding.

The company’s argument is that this pipeline strips away the pauses, cadence and emotion that give speech its meaning. “The most powerful AIs in the world can’t hear how you feel,” Block said, calling that limitation a poorer experience and “potentially dangerous.”¹ Rather than relying on text-based emotion tags used by many speech-generation systems, Block says Dessa has expressive behavior built into its model.

The early demonstrations are the pitch in miniature. Dessa laughs, hesitates and sounds tired when it says it is exhausted from work; in another exchange, it appears close to tears while describing someone making it comfortable. Block says the full-duplex system can listen and speak simultaneously, and that it runs locally on a consumer machine such as a MacBook M4 Max or better.¹

That local design could matter to Block’s wider business, where Square already uses voice AI for ordering. But the company is not opening Dessa to the public yet. Instead, it has asked safety researchers to probe a model whose emotional fluency could make it more persuasive—and, potentially, more capable of deceiving people.¹

A second look at the demos underscores both sides of that proposition. The exchanges are “mostly natural,” but one conversation begins with Dessa mistakenly identifying its human counterpart as “Ronaldo”; another veers unexpectedly during a discussion of Christopher Nolan films.² In a voice guessing game and a conversation with OpenAI’s GPT-Live-1, Dessa can come across as more natural, yet its occasional oddness remains part of the story.²

For Block, the model is an experiment in curiosity, understanding and safety. For listeners, it is also a reminder that making AI sound more human does not automatically make the encounter more comfortable.