Block releases Dessa, a full-duplex speech model that runs locally on consumer GPUs

Block, led by Jack Dorsey, announced Dessa, a full-duplex speech model trained entirely on human speech using a single NVIDIA GPU rack. The company says Dessa can run locally on consumer-grade hardware such as a MacBook M4 Max, positioning it as a step toward on-device voice AI without reliance on cloud-based pipelines.

Unlike many current systems that convert speech to text and back, Dessa is designed to handle real-time listening and speaking simultaneously, a capability Block compares to OpenAI’s GPT-Live. Block asserts that emotion and nuance are intrinsic to the model rather than driven by external text tags, which are commonly used by systems like ElevenLabs v4 or Fish Audio Drama 3.

Block argues that text-in-the-middle approaches discard rich vocal information and can produce unnatural, annoying outputs. The company released audio examples where Dessa’s voice adapts to context—sounding tired during a simulated work discussion and emotional during a role-play scenario—claiming these traits emerge from conversational data rather than engineered tags.

The model’s knowledge is derived from conversational data rather than a paired large language model, according to Block. The company also released a short video demonstrating Dessa playing a guessing game with a human, highlighting real-time responsiveness.

Block has not opened public access to Dessa. Instead, it invited safety researchers to test the model, likely to probe scenarios where a voice agent could deceive users. The company, which already uses voice AI in Square for ordering, suggests on-device models could lower costs and enable new experiences for businesses and consumers.