Most voice AI still works by turning your speech into text, thinking about the text, and turning a reply back into speech. Block thinks that approach is broken, and its new model, Dessa, is the company’s answer. It laughs, pauses, and even sounds like it’s about to cry.
Block’s research lab released Dessa as a full-duplex model, meaning it can talk and listen at the same time. OpenAI’s GPT-Live, released earlier this year, works the same way. Block says Dessa was trained entirely on human speech using a single NVIDIA GPU rack, which is a small footprint by frontier-lab standards.
The architecture matters more than the demos. Typical voice products pair a speech model with a large language model that supplies knowledge. Block says Dessa’s knowledge comes from its conversational training data instead. That could limit what it knows, and the company hasn’t said how it handles factual questions. Still, dropping the text layer is the whole point of the pitch.
Block’s argument is blunt. In its blog post, the company said current voice AI “turns your voice into text and throws away an incredible amount of information text alone can’t carry.” It added that the most powerful AIs can’t hear how you feel, which it called ridiculous and potentially dangerous.
That puts Dessa at odds with tools like ElevenLabs v4 and Fish Audio Drama 3, which use voice tags to steer emotion. Block says emotion is built into Dessa. It’s not clear whether any tag-like steering happens under the hood, and that’s a question developers will want answered.
The demos are the strongest evidence so far. In one, the bot says it’s tired from work and then sounds tired. In another, it talks about someone making it comfortable and sounds close to breaking down. A third shows a Block employee playing a game with it. These are curated clips, though, so treat them as a best case.
The local angle may be the real story. Block says Dessa runs on a consumer machine such as a MacBook M4 Max or better. That’s a high-end laptop, not a phone, but it points toward voice agents with no cloud round trip, lower latency, and lower cost per conversation.
There’s no public access yet. Block has instead asked safety researchers to test the model, most likely for cases where a voice agent could deceive the person on the other end. That caution is reasonable. A model that convincingly performs distress is a short step from one that manipulates.
For Block, the business logic is clear. Square already uses voice AI for ordering, and an in-house on-device model could power more merchant and consumer features at lower cost. So does this matter? If the quality holds outside the demo reel, yes. Until then, it’s a promising research release from a company that wants to own its voice stack.
