Block, the payments company run by Jack Dorsey, has shown off an AI voice model called Dessa that laughs, pauses, sighs and even sounds close to tears in conversation. The company says it was trained “from scratch on nothing but human conversation”, without the written text most AI models learn from.
How Dessa is different
Most voice assistants work in steps: they turn your speech into text, let a text-based AI write an answer, then read that answer out loud. Block argues that this throws away a lot of what people actually communicate, such as tone, hesitation and emotion. “The most powerful AIs in the world can’t hear how you feel,” the company wrote in a blog post quoted by TechCrunch.
Dessa skips the text step. It is what engineers call a full-duplex speech model, meaning it can listen and talk at the same time, like a person who reacts while you are still speaking. TechCrunch notes that OpenAI’s GPT-Live model works in a similar way. Block says its model’s knowledge comes from the conversations it was trained on rather than from a separate text model.
Block also says the model was trained on a single NVIDIA GPU and can run on a consumer computer: a MacBook Pro with an M4 Max chip or better.
Natural, but sometimes awkward
The demo clips are impressive and a little strange. The Verge describes one chat where Dessa mistakes the speaker for someone called “Ronaldo”, and another where a talk about Christopher Nolan films drifts into an unexpected tangent about Superman movies. In one example, Dessa says it is tired from work and actually sounds tired. In another, it responds to news of a surprise trip to Paris with what its own website calls “teary gratitude”.
Can you try it?
Not yet. There is no public access. Block says a small group of AI safety researchers will get early access first. TechCrunch suggests this is likely to test how a very human-sounding voice could be misused, for example to trick people on a phone call.
Block gives three reasons for building Dessa: curiosity, better understanding between people and AI, and safety. On safety, the team’s argument is that today’s voice AI is “nearly deaf to the emotion in our voices”, which could matter if someone in distress is talking to a machine that cannot tell.
What it means for you
Block already uses voice AI in Square to take orders for businesses. A voice model that runs on a single device could make that kind of feature cheaper and more natural. For everyone else, Dessa is a sign that AI voices are about to get much harder to tell apart from real people, which is both exciting and a reason to be more careful with unexpected calls.
1 thought on “Block’s Dessa voice AI laughs, pauses and sounds human”