Block Unveils Dessa, a Locally Running Voice Model That Sounds Human

Block released the voice model Dessa

Block, the company founded by Jack Dorsey, has unveiled a new voice model named Dessa, developed within its research laboratory. According to TechCrunch, the model operates locally and sounds surprisingly human: it expresses emotions through laughter, pauses, and even crying. The company stated that Dessa was trained entirely on human speech using a single NVIDIA GPU rack.

Full-duplex architecture without a text pipeline

Dessa is a full-duplex model: it can speak and listen simultaneously, similar to the GPT-Live model released earlier this year by OpenAI. Typically, speech models are paired with large language models for knowledge retrieval, but Block claims that Dessa's knowledge comes from its conversational data. According to the company, existing models sound unnatural because their pipelines include text that controls speech generation; models such as ElevenLabs v4 or Fish Audio Drama 3use voice tags to convey emotions. Block asserts that emotions are embedded directly into its model, though TechCrunch notes it remains unclear whether these emotions are controlled by specific tags.

Demonstrations and criticism of current voice AI

In its blog, Block stated: 'Interacting with current voice AI can simply be terrible. It converts your voice into text and discards a vast amount of information that text alone cannot convey. The most powerful AIs in the world cannot hear how you feel.' The company's researchers published several examples of conversations with Dessa, where the model used different voices, intonations, and emotions: in one dialogue, the bot said it was tired from work and sounded exhausted; in another, it nearly broke down emotionally; in a third a Block employee played a guessing game with the assistant, which the model solved quickly. Jack Dorsey wrote on Xthat 'talking to her feels like talking to a human… full of emotions, nuances, pauses, and natural rhythm'. Researcher Ragavan Thurairatnam called Dessa 'the GPT-3 moment for speech'.

Local deployment and access

According to the company, the model can run locally on consumer devices such as a MacBook M4 Max or better. Public access to Dessa is currently unavailable: Block has invited security researchers to test the model — likely to examine scenarios where a voice agent might deceive a conversation partner.

Implications for Block's business

Block operates in the financial sector and sells solutions to small businesses; Square already uses voice AI for order processing. An on-device proprietary model will enable the company to develop more business and consumer scenarios at lower costs. Editorial note: The announcement is notable because Dessa was trained without an intermediate text layer, but Block's key claims regarding 'embedded' emotions and sound quality are currently supported only by the company's own demonstrations, while independent verification is limited due to restricted access. It is also unknown how the model scales beyond a single GPU rack and when public access will become available.