Block Unveils Dessa, an AI That Learns Speech Without Text

Jack Dorsey’s Block has unveiled Dessa, a speech-to-speech AI model designed to interact more like a natural human conversation.
Unlike traditional voice assistants that convert speech into text before generating a response, Dessa was trained directly from audio without relying on text as an intermediate layer.
That allows the model to learn more than just words.
It can interpret elements such as tone, pauses, sarcasm, sighs and emotional cues while producing similarly expressive speech in response.
Dessa can also listen and speak at the same time, making interruptions and back-and-forth exchanges feel closer to a real conversation.
The model can reportedly laugh mid-sentence and respond to subtle vocal signals that are often lost when speech is first converted into text.
For $XYZ, the technology highlights a broader shift toward native audio AI rather than voice interfaces built on top of text-based models.
If these systems scale effectively, they could significantly expand the use of AI in customer support, assistants, commerce and real-time communication.
The bigger implication is that AI interfaces may increasingly move beyond typing and toward continuous, emotionally aware voice interaction.