Block's New AI Model Learns Language Only From Human Voices

Block, the company founded by Jack Dorsey, has released an artificial intelligence model named Dessa that processes audio inputs directly without converting them to text. The system was trained exclusively on human conversation, a method the developers say mirrors how infants acquire language skills before they learn to read. Block published recordings of the model interacting with users to demonstrate its capabilities, showing both natural exchanges and moments of apparent confusion.

Ragavan Thurairatnam, who leads the project at Block, shared one sample interaction where the model initially misidentified the speaker as a person named Ronaldo. Other examples on the company's website show Dessa discussing Christopher Nolan films before drifting into a tangent about Bryan Singer's Superman movies. In another clip, a user tells the model about a surprise trip to Paris, and Dessa responds with what the website describes as teary gratitude. A separate demonstration pairs Dessa with OpenAI's GPT-Live-1 voice model, resulting in a more fluid conversation.

The team states that Dessa is a true speech-to-speech model capable of running locally on a MacBook Pro M4 Max or better hardware. This local processing capability allows the system to operate without sending data to cloud servers. The developers cite curiosity, improved human-AI understanding, and safety as their primary motivations for building the system without text data. They compare the approach to baby development, noting that children learn to talk without seeing written words. The team acknowledges that training a model this way is difficult but confirms it is possible.

The developers argue that current voice assistants underutilize the potential of spoken language. They claim that existing systems flatten speech into text and fail to match the timing of natural human conversation. The safety argument for the voice-only approach focuses on emotional recognition. The team notes that AI systems are gaining power while remaining nearly deaf to the emotion in human voices. They describe a scenario where a person in serious pain interacts with a voice AI that continues the conversation unaware of the distress. Dessa is designed to address this gap by analyzing the audio signal directly.

Block plans to provide early access to the model for a small group of AI safety researchers. The announcement did not include a timeline for a wider public release. The company's focus on local processing and emotion-aware speech aims to create a more responsive interface for users who communicate primarily by voice.