Blocks latest AI model Dessa was trained entirely on human speech — and it shows

Dessa, an AI model developed by Block, the entity led by Jack Dorsey, is reportedly the world's first full-duplex speech model. This means it can listen and speak at the same time, similar to how humans converse naturally. The model's unique feature lies in its training - it was trained exclusively on human speech, resulting in a natural-sounding dialogue.

Unlike traditional voice assistants like Siri, which often struggle with basic commands and processing times, Dessa aims to provide a more human-like interaction. The model's proficiency is demonstrated in several audio samples. In one, Dessa engages in a conversation about Christopher Nolan's movies with a human, responding with knowledge that comes across as genuine and unforced.

In another, Dessa's reaction to an unexpected prompt - expressing excitement and enthusiastic responses - mirrors typical human reactions, including verbal expressions and emotional responses. Perhaps most impressively, Dessa can discern sarcasm, a complex task that requires an understanding of tone and its relationship to the content of speech.

This ability highlights Dessa's advanced comprehension of human conversation. Block's motivation for creating Dessa is clear: they believe speech is the most natural way to interact with any intelligence. Traditional voice AI, they argue, fails to capture the subtleties of human speech and the emotional undertones that accompany it.

Dessa, they claim, is a significant step towards overcoming this challenge, making AI more in tune with human communication and understanding.