Latest
Category 2 Hurricane Isaias nears landfall, leading to production shut-insWhite House says the US is lifting tariffs on some Canadian goodsBodycam: Cop Was Not in Danger When He Fired Shots That Killed 1-Year-Old Kohen WileyCollins says certain Canadian goods now tariff-exemptDemocrats are warning their party not to get too bullish on House gainsIndiana police officer arrested for using Flock cameras to search wifeUkraine: Starlink still isn’t working after Trump restoration promiseCategory 2 Hurricane Isaias nears landfall, leading to production shut-insWhite House says the US is lifting tariffs on some Canadian goodsBodycam: Cop Was Not in Danger When He Fired Shots That Killed 1-Year-Old Kohen WileyCollins says certain Canadian goods now tariff-exemptDemocrats are warning their party not to get too bullish on House gainsIndiana police officer arrested for using Flock cameras to search wifeUkraine: Starlink still isn’t working after Trump restoration promise
← Signal: Tech & AI

Block announces Dessa, an AI model trained on human speech rather than text

Confirmed1 source · Oct 23, 2026

Jack Dorsey's Block has unveiled Dessa, a speech-to-speech AI model trained exclusively on human conversation that can run locally on high-end MacBooks.

Block announces Dessa, an AI model trained on human speech rather than text
Image via The Verge

What happened

Block announced an AI model called Dessa that was trained "from scratch on nothing but human conversation" rather than text. The model is a "true speech-to-speech model" capable of running locally on a MacBook Pro M4 Max or better. Block demonstrated Dessa through published conversation clips, including exchanges about films, a trip to Paris, and a voice guessing game, though some interactions exhibited awkwardness—such as initially mistaking a speaker for someone named "Ronaldo." The Dessa team stated the model was created out of "curiosity," to improve human-AI understanding, and for safety reasons, arguing that existing voice AI flattens speech into text and fails to detect emotional tone.

Context

Dessa represents an alternative approach to voice AI by bypassing text conversion and operating directly on speech, which the team frames as more natural for human-machine interaction. The stated safety rationale—that current voice AI cannot perceive emotion or distress—suggests an underlying concern that speech-based systems missing human emotional cues could fail in situations requiring contextual sensitivity. Early access is being restricted to a small group of AI safety researchers, indicating the developers are treating this as an early-stage project requiring specialized feedback rather than a consumer-ready product.