On October 24, 2026, Block, the company under Jack Dorsey, officially released the all-new voice model Dessa. The model is trained entirely on human speech data, supports local on-device deployment, and features remarkably lifelike pronunciation along with rich emotional simulation capabilities such as laughter, pauses, and crying.
As a full-duplex speech-to-speech model, Dessa can listen and talk simultaneously, and its knowledge base comes directly from conversational data, with no additional reliance on large language models for knowledge retrieval.
Traditional voice models mostly use text as their generation input, often sounding stiff because they lack the subtle context behind the words. By contrast, Dessa has built-in natural emotion-capturing ability, effectively solving the pain point of existing voice AI easily losing emotional information and subtle tonal nuances during interaction.
In terms of hardware compatibility, the model can run efficiently on consumer-grade devices such as a MacBook M4Max or a single NVIDIA GPU rack. At present, Block has not yet opened the model fully to the public, instead inviting security researchers in the first batch to participate in testing to evaluate its performance in different scenarios.
As a fintech company focused on small business solutions, Block has previously experimented with applying voice AI to order processing through Square. With the arrival of Dessa, which can run locally on end devices and offers highly lifelike interaction, businesses and end users can expect a more immersive and natural intelligent interaction experience at lower cost, marking an acceleration of voice AI's evolution toward a new stage of native voice interaction.