A new AI model called Dessa is turning heads in the voice AI world, and its creators are not being shy about it. They describe it as a full-duplex speech-to-speech model, meaning it can listen and talk at the same time, just like a person does in a real conversation. According to the team, it was trained solely on raw human conversation audio, with no text and no pretrained models behind it. They say this lets Dessa handle interruptions, pauses, emotion, sarcasm and little "uh-huh" style reactions that most voice assistants still fumble. The developers call it a "GPT-3 moment" for speech, a bold comparison that invites plenty of scrutiny.

The demo video shows a fast, back-and-forth chat between a person and Dessa, complete with casual slang, laughter, interruptions, empathy and quick replies like "yeah" and "got it." The exchange sounds loose and natural rather than scripted, which is exactly what the creators are pitching. Dessa comes from an AI lab backed by Jack Dorsey's Block, and the team says it can run locally on consumer hardware, including a MacBook Pro M4 Max. Those are the developers' own claims, and independent testing will show whether the model holds up outside a polished demo. Still, if it performs as shown, it raises a big question for everyone building voice assistants: do they really need text as a middleman at all?