Block has unveiled Dessa, a new AI model built to have realistic spoken conversations with people. Unlike many assistants that are primarily trained around text and then adapted for voice, Dessa was reportedly trained from scratch on human conversation.
The result is a speech-first system designed for full-duplex interaction, allowing more natural back-and-forth exchanges. Early shared demos show conversations that can flow surprisingly well, even if the model sometimes still sounds awkward or makes odd mistakes.
Why this matters
Voice is one of the most intuitive ways for people to interact with technology, especially for users who may find typing difficult, slow, or inaccessible. A model designed around spoken conversation from the beginning could help create AI assistants that feel more responsive, approachable, and useful in real-world settings.
- More natural interaction: Speech-first training may improve timing, turn-taking, and conversational rhythm.
- Accessibility potential: Better voice AI can support people who rely on spoken interfaces.
- Real-world promise: Even imperfect demos suggest progress toward AI that communicates less like a chatbot and more like a conversation partner.