Block's new AI model trained on 100% human audio

Jack Dorsey’s company Block released a new artificial intelligence model named Dessa today. The system was trained exclusively on human conversation. It does not process text during training or operation. Block described the release as a true speech-to-speech model. All of the model’s knowledge came from listening to people talk.

The project team, led by Ragavan Thurairatnam, shared sample recordings to demonstrate the system. In one clip, the model initially mistakes the speaker for a person named Ronaldo. The conversation then proceeds with minor awkwardness. Another sample features a discussion about Christopher Nolan films that drifts toward Bryan Singer’s Superman movies. A third clip pairs Dessa with OpenAI’s GPT-Live-1 voice model. In that specific exchange, the model sounds more natural to the ear.

Block has operated the project for several months. The team began with a goal to mimic how infants learn language. They noted that babies learn to speak without seeing written words. The developers wanted to test if a machine could achieve the same result. They confirmed that it can, though they described the process as very difficult. The project started with a focus on curiosity and fun.

The developers argue that current voice assistants limit the potential of speech. They stated that existing systems flatten spoken words into text. This conversion causes the AI to lose track of human timing. Block believes speech is the most natural interface for any intelligence. The new model attempts to remove the text layer entirely. This change is designed to improve the speed and feel of the interaction.

Safety is the third motivation for the project. The team says modern AI systems are becoming more independent. However, they remain nearly deaf to the emotion in human voices. Block provided a scenario where a person is in serious pain. A standard voice AI might carry on regardless because it cannot hear the distress. Dessa aims to fix this by processing the audio directly.

The model runs locally on a MacBook Pro M4 Max or better hardware. It does not require cloud processing for basic operation. Block plans to give early access to a small group of AI safety researchers. The announcement did not include a date for a wider public release. The team said the next step is to let those researchers test the system in controlled environments.