Block unveils Dessa voice AI model trained solely on conversations — The Verge

Jack Dorsey’s company Block has announced Dessa, an artificial intelligence model for voice communication that, according to its developers, was trained from scratch exclusively on human speech. The team calls it a full-duplex speech-to-speech model, meaning a system that works directly with voice.

Examples of voice dialogues

As The Verge reports, Block released recordings of Dessa’s conversations with people. In some of them, the model conducts a mostly natural dialogue, although in one case it mistakenly addressed its interlocutor as Ronaldo at the beginning. The company also showed Dessa playing a voice-based guessing game.

The project website also features other examples: a discussion of Christopher Nolan’s films, a conversation about an unexpected trip to Paris, and a dialogue between Dessa and OpenAI’s GPT-Live-1 voice model. During the conversation about cinema, Dessa unexpectedly shifted to Bryan Singer’s Superman films.

More current news is available on the UA.News Telegram channel Telegram.

Local deployment and early access

The Dessa team says the model is a “true” speech-to-speech system and that all of its knowledge comes from human speech. According to the developers, it can be run locally on a MacBook Pro with an M4 Max chip or a more powerful device.

The team explained that it created Dessa out of interest in whether a model could learn to speak without written text, a desire to improve mutual understanding between people and AI, as well as safety considerations. The developers believe that modern voice systems often convert speech into text and do not sufficiently account for emotions and the pace of conversation. The team plans to provide early access to the model to a small group of artificial intelligence safety researchers.