Block Debuts 'Awkward' AI Voice Model

Block, the company led by Jack Dorsey, has unveiled an AI model named Dessa that is trained exclusively on human conversation rather than written text. According to reporting from The Verge, the model is a "true speech-to-speech model" capable of running locally on hardware such as a MacBook Pro M4 Max.

To demonstrate the model's capabilities, Block shared several voice interactions. These include a guessing game, a discussion regarding Christopher Nolan's films, and a scenario where a user describes a trip to Paris, prompting a response described by the website as "teary gratitude." The Verge notes that some of these interactions can appear "uncanny" or "awkward," citing one example where the model mistakenly identifies a speaker as "Ronaldo."

Ragavan Thurairatnam, the lead for Dessa, shared some of these conversations to highlight the model's natural back-and-forth flow. The Dessa team stated that the project was driven by curiosity, a desire for better human-AI understanding, and safety. Specifically, the team argued that current voice AI fails by flattening speech into text and ignoring emotional cues—a gap they believe creates a safety risk in scenarios where a user is in pain.

While the model is not yet widely available, the Dessa team indicated that early access will be granted to a small group of AI safety researchers.