The Block Dessa AI voice model aims to make conversations with artificial intelligence sound more natural by learning directly from human speech. Jack Dorsey’s financial technology company introduced Dessa on Oct. 23, demonstrating an assistant that can laugh, pause, and express emotion during live conversations.
Dorsey introduced the technology in a post on X, describing it as a full-duplex speech model. Unlike conventional voice systems, Dessa can listen and speak at the same time, letting users interrupt or respond during conversations.
Block says the model was trained entirely on human conversations without text or pretrained language models. According to the company, researchers trained it on a single Nvidia GB200 NVL72 system.
How the Block Dessa AI Voice Model Works
Traditional voice assistants often convert speech into text before generating responses. That process can discard emotional cues, including hesitation, sarcasm, changes in tone and laughter.
Block developed Dessa to process spoken information directly. The company says this allows the system to recognize conversational timing and produce more expressive responses without relying on text-based instructions.
Other voice-generation systems, including Fish Audio’s Drama 3, support emotional expression through text instructions and speech tags. Dessa takes a different approach by learning conversational behavior from audio training data.
The company published several recorded demonstrations showing Dessa speaking with different voices, rhythms and emotional expressions. In one example, the model sounds tired while discussing work. Another features an emotional conversation in which its voice appears close to breaking.
Block also demonstrated Dessa participating in interactive games and conversations with OpenAI’s GPT-Live voice model. TechAmerica.ai previously covered OpenAI’s GPT-Live technology, which similarly supports simultaneous listening and speaking.
Dessa Runs Locally but Is Not Yet Publicly Available
Block says Dessa can operate locally on a MacBook Pro equipped with an M4 Max processor or more powerful hardware. This could reduce dependence on cloud-based processing for supported voice applications.
However, Dessa is not currently available to the general public. The company is inviting selected AI safety researchers to evaluate the model and identify potential risks before wider access.
Those concerns are especially relevant for increasingly realistic synthetic voices, which can enable impersonation and deception. Independent testing will also be important for assessing the model’s reliability beyond company demonstrations.
Block’s existing business includes Square, which already uses voice AI for restaurant ordering. More capable local voice models could eventually support additional customer service and business applications, although Block has not announced commercial deployment plans for Dessa.
The release adds another approach to the growing conversational AI market. Companies including ElevenLabs are developing more expressive speech systems, with ElevenLabs’ v4 models emphasizing greater control over voice generation.
For now, Dessa remains a research project. Block has demonstrated its conversational abilities, but broader availability, independent performance results, and commercial applications remain unestablished.
