Block deploys Dessa, a speech-only AI that skips text training

Block, the financial technology company founded by Jack Dorsey, has released a new artificial intelligence model named Dessa. The system is designed to process and generate language using only audio input, bypassing the standard conversion of speech to text. Block describes the model as being trained "from scratch on nothing but human conversation." This approach distinguishes Dessa from current voice assistants that rely on intermediate text processing.

Ragavan Thurairatnam, who leads the project, shared a demonstration recording of the model in conversation. In that sample, the dialogue appears largely natural, though the model initially misidentifies the speaker as a person named Ronaldo. The company has hosted additional audio examples on the Dessa website. In one clip, the discussion moves from Christopher Nolan films to a tangent about Bryan Singer's Superman movies. In another, a user describes a surprise trip to Paris, prompting what the site describes as teary gratitude from the model. A third example pairs Dessa with OpenAI's GPT-Live-1 voice model, resulting in more natural-sounding output.

The development team states that Dessa functions as a true speech-to-speech model. All data used for training came from human speech. The model is capable of running locally on a MacBook Pro equipped with an M4 Max chip or better. This local processing capability allows the system to operate without relying entirely on cloud infrastructure. Block positions this hardware requirement as a key feature for users seeking private, low-latency interactions.

The team outlines three primary motivations for the project: curiosity, improved human-AI understanding, and safety. They compare the model's training to human child development, noting that babies learn to speak without seeing written words. The developers state that this experiment proved a model could achieve the same result. They argue that existing voice AIs flatten spoken input into text, which disrupts natural timing. The safety argument focuses on emotional recognition. The team claims current systems are "nearly deaf to the emotion in our voices." They suggest Dessa represents a step toward fixing this gap by processing emotional cues directly from audio.

Block plans to provide early access to a small group of AI safety researchers. The announcement did not include a timeline for a wider public release. The company did not specify when general availability would begin or what hardware requirements would apply to future updates. The current release remains focused on demonstrating the viability of speech-only training architectures.