Block has introduced Dessa, an AI model built to converse through speech. Its demonstrations show a system that can keep a conversation moving, but also produce surprising or awkward responses—an early glimpse of what voice-first AI can do and where it still struggles.
A conversational model with rough edges
Block says Dessa was trained “from scratch on nothing but human conversation.” The company has shared recordings of people speaking with the model, including a mostly natural back-and-forth in which Dessa initially mistakes the speaker for someone named “Ronaldo.”
Other examples cover a guessing game and a discussion of Christopher Nolan’s films. In the latter, Dessa unexpectedly veers into Bryan Singer’s Superman films. These detours make the exchanges feel less like polished demonstrations and more like conversations with an unpredictable partner.
The Dessa website also features a person describing a surprise trip to Paris. The model responds in a way the site characterizes as sounding like “teary gratitude.” Another clip pairs Dessa with OpenAI’s GPT-Live-1 voice model, with Dessa coming across as more natural in that exchange.
Why build AI around speech?
The team describes Dessa as a “true speech-to-speech model,” saying that everything it knows came from human speech. That design is meant to let the system work with spoken conversation directly, rather than treating speech only as a route to text.
Block’s team says it began with a question about how people learn to talk: babies do so without seeing written words, and the researchers wanted to find out whether a model could learn in a similar way. They say it can, while emphasizing that doing so is difficult.
The team also argues that voice is a natural way for people to interact with intelligence, but that current voice AI can flatten speech into text and fail to match human timing. Dessa is an attempt to explore a different approach to that interaction.
Conversation and safety
Block gives three motivations for the work: “curiosity (aka fun), better understanding between humans and AI, and safety.” The safety case centers on a limitation the team sees in voice systems: AI can miss emotion conveyed through a person’s voice.
The team illustrates the concern with someone in serious pain speaking to an AI that does not recognize the distress and simply continues. In that scenario, the system’s failure is not just a clumsy answer; it is a failure to respond to an important human signal. Block presents Dessa as a step toward addressing that gap.
The demonstrations also show why that goal is challenging. A model can sound engaged and respond with apparent feeling, yet still misidentify a speaker or wander onto an unrelated topic. Natural-sounding conversation and reliable understanding are separate demands, and the clips offer evidence of both promise and unevenness.
Local operation and early access
According to the Dessa website, the model can run locally on a MacBook Pro M4 Max “or better.” The team says it will provide early access to “a small group of AI safety researchers.”
That limited access connects the project’s technical experiment to its stated safety purpose. The material shared so far presents Dessa as a research effort into speech-based interaction: one that can produce convincing moments, but also reveals how much context, timing and emotional understanding a voice model must handle to converse well.