Block introduces Dessa, a voice model trained on human conversation
Block has introduced Dessa, a direct speech-to-speech research model that the company says learned entirely from human conversation. It is designed to listen while speaking and retain timing and emotion that text conversion can lose. Published demonstrations accompany a public waitlist and planned early access for a small group of safety researchers.
Artificial Intelligence··Night
Dessa learns from speech without a text stage
Block, the technology company cofounded by Jack Dorsey, introduced its Dessa speech model on October 23. The company says it trained the model from scratch entirely on human conversation, without text or pretrained models. Dessa takes speech directly as input and generates speech as its response. Recordings released with the announcement show exchanges with people and with another AI voice model.[1], [2]
The design is intended to let the model listen while speaking, learning when to interrupt or offer a brief acknowledgment. Block argues that converting speech into text removes information carried by emphasis, pauses, laughter and emotion. Its stated safety concern includes a voice system continuing to talk when the person addressing it is in serious pain. These capabilities and goals come from Block’s description of the project.[2], [1]
Released conversations show games and unexpected turns
The demonstrations include a spoken guessing game and help with remembering a film title. In a conversation shared by Thurairatnam, the model initially mistakes the speaker for someone called Ronaldo. Another exchange moves from Christopher Nolan’s films to Bryan Singer’s Superman films. Block also published an example in which a person describes a surprise trip to Paris; the company characterizes Dessa’s response as sounding emotionally grateful.[1]
A separate recording pairs Dessa with OpenAI’s GPT-Live-1. Block says the other model’s developer has neither reviewed nor endorsed Dessa and is not affiliated with the project. The company says Dessa chooses what to say, when to speak and how to deliver its response in the demonstrations, with subtitles added afterward. Other examples include responding to a request to speak louder and reacting to an anniversary announcement.[2], [1]
Safety researchers receive early access
Block has opened a public waitlist and plans early access for a small group of AI safety researchers to uncover risks the team has not anticipated. The research lab was founded by former members of Toronto deep-learning startup Dessa and reports directly to Dorsey. The team names curiosity, better understanding between people and AI, and safety as its reasons for pursuing learning directly from sound.[2], [1]
