Block, the company led by Jack Dorsey, has released a voice model called Dessa from its research lab. It can laugh, pause and even sound like it is crying, and it runs locally.
What makes it different
Most voice assistants work as a chain: speech becomes text, a language model handles the text, and the reply is read aloud. Block argues that the first step throws away everything text cannot carry, such as tone, fatigue and hesitation. Dessa skips that layer. Emotion is built into the model itself, and its knowledge comes from the conversations it was trained on. Training used only real human speech and a single rack of Nvidia GPUs.
The model is full duplex: it talks and listens at the same time, so you can interrupt it the way you would a person. Block says it will run on consumer hardware, from a MacBook M4 Max upward.
What was shown
Block staff posted recordings where the agent switches voices and pace. In one, it says it is tired after work and actually sounds tired. One researcher called the release a "GPT-3 moment for speech."
Caveats
There is no public access yet, and nearly all quality claims come from the company and its employees. Block has asked outside safety researchers to test the model, including whether it could be used to deceive people. For Block's business this could mean cheap on-device voice ordering and support, and Square already uses voice AI to take orders.