Jack Dorsey’s Block Unveils Dessa, an AI Trained Only on Speech

Ender Öztürk

October 26, 2026

Block, the financial technology company led by Twitter co-founder Jack Dorsey, has introduced Dessa, a pioneering artificial intelligence model designed to engage in natural, real-time human conversations. Unlike traditional AI systems that rely on massive datasets of text to function, Dessa was developed through an innovative approach where it was trained entirely on raw audio data. By bypassing the need for text conversion, this model learns the nuances of human speech, including emotional depth and rhythmic pacing, simply by listening. This development represents a significant shift in how voice-based AI architectures are constructed and deployed in the modern tech landscape.

The Architecture Eliminates Traditional Text Processing

Most modern voice assistants operate through a multi-stage process that converts speech to text, processes the information via a large language model, and then converts the resulting text back into synthetic speech. This workflow frequently strips away the subtle emotional cues, tonal variations, and natural pauses that define organic human communication. Dessa changes this dynamic by utilizing an end-to-end speech-to-speech architecture.

By removing the textual intermediary, the model processes acoustic input directly, allowing it to capture the unique prosody of human speech. This methodology mirrors how humans acquire language through auditory exposure rather than reading. Consequently, the model exhibits a more authentic conversational flow, maintaining the cadence and emotional weight that text-based models often struggle to replicate accurately during live interactions.

Full-Duplex Technology Enables Real-Time Interactivity

A standout technical achievement in the Dessa project is its implementation of full-duplex communication. In standard voice interfaces, users are often forced to wait for the system to stop speaking before they can interject. Dessa, however, is designed to listen while it speaks, creating a genuine back-and-forth dialogue experience.

This feature allows the AI to react immediately if a user interrupts, provide short verbal feedback, and adjust its response rhythm based on the flow of the conversation. While early demonstrations show the model successfully engaging in games and casual discussions, researchers have noted that the system occasionally drifts from the topic or misidentifies the interlocutor, indicating that the technology remains in an experimental phase.

Hardware Requirements Define Local Processing Capabilities

Block has confirmed that Dessa is capable of running locally on hardware such as the M4 Max-powered MacBook Pro. This local execution capability ensures that the AI can perform inference tasks faster than real-time without needing constant cloud connectivity. Despite these promising benchmarks, the company has not yet provided comprehensive data regarding the model’s size or its performance latency across a wider range of hardware configurations.

As the project remains in a research preview stage, the team is prioritizing safety and stability. By restricting access to a specialized group of researchers, Block aims to identify potential vulnerabilities before considering a broader release. This cautious strategy reflects the broader industry trend of prioritizing ethical AI development in speech-to-speech technologies.

Do you believe that speech-to-speech models like Dessa will eventually render traditional text-based voice assistants obsolete, or do you prefer the precision of text-processed responses? Share your thoughts in the comments section below.