Block launches Dessa, a voice-to-voice model trained on speech

Block, the company led by Jack Dorsey, has introduced Dessa, an AI model claimed to be trained solely on human conversations. The team emphasizes security motivations and plans to offer early access exclusively to researchers. They also assert that the system can operate locally on a MacBook Pro M4 Max or better. Published examples of exchanges show a tone that is sometimes natural and sometimes awkward.

Restricted Access and Security Highlighted by the Team

The Dessa team plans to open early access to a small group of AI security researchers. They explicitly cite security as one of the reasons for developing the model, believing that AI is gaining power and independence while remaining almost deaf to the emotions in voices. They illustrate the risk with a scenario of intense pain that the voice AI would not perceive, where it would continue to speak. Dessa is presented by the team as a step towards addressing this issue. Furthermore, the team claims that the model can function locally on a MacBook Pro M4 Max or better.

A Speech-to-Speech Model Derived Exclusively from Oral Data

According to the team, Dessa is a true speech-to-speech system, with knowledge derived solely from human speech. They assert that speech represents the most intuitive interface for any intelligence but believe that current voice AI systems reduce spoken words to text and do not respect the rhythm inherent in human communication. To explain their approach, the team notes that children learn to speak without ever being exposed to written language and states that they wanted to verify if a model could replicate this process, an ambition described as achievable but very complex.

Public Demonstrations with Uneven Results

To illustrate Dessa, Block has shared several exchanges with humans. Ragavan Thurairatnam, who leads the project, broadcast a dialogue where the model initially confuses the interlocutor with a person named Ronaldo. A video also shows it playing a vocal guessing game. The site features other interactions, including a discussion about Christopher Nolan films where Dessa veers off into Bryan Singer's Superman, and a scene where someone claims to have taken Dessa to Paris, described as eliciting heartfelt gratitude. The site also offers a conversation with OpenAI's GPT-Live-1, where Dessa appears more natural. Audio clips of these exchanges are available in video format.