Block Unveils Dessa, a Pure Speech-to-Speech AI Trained Only on Human Conversation

Jack Dorsey’s technology company Block has announced Dessa, an artificial intelligence model trained entirely on human speech rather than written text. Led by Ragavan Thurairatnam, the development team built the system "from scratch on nothing but human conversation" in an effort to create a more natural and emotionally responsive voice interface.

A Pure Speech-to-Speech Approach

Unlike traditional voice assistants that convert spoken words into text, process the text, and convert the output back into synthesized speech, Dessa is a true speech-to-speech model. Everything the system knows was learned directly from human speech. The team compared this training philosophy to childhood development, noting that babies learn to speak without ever seeing written text.

On the technical side, the model is capable of running locally on high-end consumer hardware, requiring an Apple MacBook Pro with an M4 Max chip or better.

Uncanny Interactions and Emotional Nuance

Demonstrations published on the Dessa website highlight both the realism and the occasional awkwardness of the model. In audio clips shared by the team, Dessa engages in back-and-forth guessing games and discusses cinema, including a conversation about filmmaker Christopher Nolan where the model goes on an unexpected tangent about Bryan Singer’s Superman movies.

In another shared interaction, when a user tells Dessa about taking it on a surprise trip to Paris, the AI responds with what the team describes as teary gratitude. However, the model still exhibits unpredictable behavior—in one conversation led by Thurairatnam, Dessa initially mistook the speaker for someone named "Ronaldo."

The website also features a dialogue between Dessa and OpenAI’s GPT-Live-1 voice model, in which Dessa demonstrates a distinctly more natural conversational cadence.

Addressing Emotion and Safety in Voice AI

Block's Dessa team cited three main motivations behind the project: curiosity, improving human-AI understanding, and safety.

According to the developers, existing voice AI tools flatten human dialogue into text, losing crucial context regarding speech timing and vocal emotion. They pointed out that current models remain largely blind to emotional signals, creating safety risks when a user might be in distress or pain that a standard text-based voice system cannot detect.

Block announced that it will initially provide early access to Dessa to a small group of AI safety researchers to conduct further evaluations.