Dessa: Jack Dorsey's Block Unveils AI That Learns from Sound, Not Text
Jack Dorsey's Block unveils Dessa, a groundbreaking AI voice model trained entirely on human conversation without using text.
In a bold departure from how most artificial intelligence is built, Jack Dorsey-led fintech giant Block has introduced Block Dessa AI voice model, a revolutionary speech-to-speech system. Unlike the industry giants that rely on massive libraries of written text, Dessa was trained exclusively on actual human conversations, allowing it to master the art of talking by simply listening.
Breaking the Text Barrier
Most conventional voice AI systems operate through a multi-step process: they convert speech into text, process that text, and then convert the result back into a synthetic voice. Block claims that this process strips away the "soul" of communication—the emotional cues, the tone, and the subtle variations that make human speech authentic.
Dessa bypasses this entirely. As a full-duplex speech-to-speech model, it processes audio directly. This means it can listen and speak at the same time, preserving the nuance of human interaction. According to the company, the model was built from scratch using nothing but sounds, without the help of any pretrained models or text-based data.
Emotion, Sarcasm, and Laughter
Because it learned from real people, Dessa doesn't just deliver information; it delivers emotion. The AI is capable of recognizing and expressing a wide spectrum of human feelings, including:
- Sarcasm and Humour: The model can react naturally to a joke, even laughing mid-sentence.
- Emotional Depth: It can convey frustration, sadness, and joy through its voice.
- Natural Cadence: It utilizes pauses and rhythms that mimic a real person, making the interaction feel authentic rather than robotic.
The company further noted that Dessa can independently decide what to say, when to jump into a conversation, and how to style its voice to fit the mood of the interaction.
Efficiency Over Bulk
Training an AI on speech is notoriously difficult because spoken words contain significantly more data than written text and vary every time they are uttered. However, Block highlights a surprising efficiency in Dessa's learning process.
While Large Language Models (LLMs) require staggering amounts of text to function—learning everything from conversational English to coding languages like C++ and Angular—Dessa learned to communicate using hundreds of times fewer words. Remarkably, Block stated they achieved this training using only a single NVIDIA GB200 NVL72.
Prioritizing AI Safety
Acknowledging that a model trained entirely on sound is "new territory," Block is not releasing the technology to the general public immediately. To ensure the system is safe and to identify unforeseen risks, the company is granting early access to a select group of AI safety researchers.
"We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built," the company stated.