Block unveils Dessa AI trained entirely on human conversation

Jack Dorsey’s company Block has unveiled Dessa, a voice-based artificial intelligence model trained entirely on human conversation rather than written text. The model runs locally on specific Apple hardware and utilizes full-duplex technology to converse, laugh, and interrupt in real time without accessing the internet.

Block Trains Audio Model on NVIDIA System

Block trained Dessa from scratch using recorded human speech on a single NVIDIA GB200 NVL72 system, which starts at $20,000. According to the company, this marks the first time an AI model trained exclusively on audio has achieved realistic, real-time conversational capabilities without reading a single word of text. The resulting model runs locally on a Mac. However, it requires a MacBook Pro equipped with an M4 processor or better to function.

Real-Time Conversational Mechanics

Dessa uses full-duplex technology, allowing the system to listen and speak simultaneously during an exchange. In public demonstrations published by Block, the model reacts spontaneously to humor, laughs mid-sentence, and utters short acknowledgments while the other person is still talking. The AI perceives sarcasm, frustration, and emotional shifts in a speaker’s voice, adjusts its volume when asked to speak louder, and fills in forgotten words or film titles when a partner pauses mid-sentence.

Frequently Asked Questions

• How was Dessa trained if it never read text?
  Block trained the model from scratch using audio recordings of human conversations, allowing the system to learn language, knowledge, and conversational skills entirely through listening.
• What equipment is required to run Dessa locally?
  Running the model locally requires a MacBook Pro with an M4 processor or a superior Apple M-series chip.
• Which company developed Dessa?
  A small research team within Jack Dorsey’s company Block developed the model, which Dorsey presented on social media.