Block's Dessa runs locally on M4 Macs after training on pure human audio

Block has launched a new artificial intelligence model called Dessa that was built entirely from human voice data. The company states the system was trained from scratch using nothing but conversation, skipping the traditional reliance on written text.

The model is designed to operate locally on consumer hardware. Block specifies that Dessa can run on a MacBook Pro equipped with an M4 Max chip or a more powerful processor. This local processing capability allows the AI to function without sending user audio to external servers.

Ragavan Thurairatnam leads the project team at Block. He released sample audio clips to demonstrate the model's performance. In one recording, the system engages in a dialogue that appears mostly fluent, though it initially mistakes the human speaker for a person named Ronaldo. Other examples on the project website show the AI discussing film directors and reacting to a user's travel story to Paris with what the team describes as teary gratitude.

The developers characterize Dessa as a true speech-to-speech system. They argue that current voice assistants flatten human speech into text, which causes a loss of nuance and timing. The team notes that babies learn to talk without seeing written words, and they aimed to replicate that learning process for a machine. They acknowledge the technical difficulty of this approach but confirm the model successfully acquired language skills solely from audio.

The team cites three primary motivations for building Dessa: curiosity, improved human-AI understanding, and safety. They argue that existing AI systems often ignore the emotional cues in a user's voice. The developers offer a scenario where a person in pain interacts with a standard voice assistant that fails to detect the distress. They describe Dessa as a step toward addressing that gap by processing the emotional weight of speech directly.

Block will provide early access to a small group of AI safety researchers. The announcement did not include a timeline for a wider public release or specific pricing for the model.