Jack Dorsey’s Block has unveiled a new voice model from its research lab that runs locally and mimics human speech with laughs, pauses and even cries. The model, named Dessa, was trained entirely on human recordings using a single Nvidia GPU rack. Dessa operates as a full‑duplex system, meaning it can speak and listen at the same time, similar to OpenAI’s GPT‑Live model released earlier this year. While most speech systems rely on a large language model for knowledge retrieval, Block says Dessa’s knowledge comes from the conversational data it was trained on.
“Voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full‑duplex speech model ever built,” the company wrote in a blog post. “Talking with it feels like talking with a person, full of emotion, nuance, pauses, and a natural cadence.” The post added that current voice AIs often sound artificial because they convert speech to text before generating audio, losing subtle cues that convey feeling. Block argues that Dessa embeds emotion directly in the model rather than relying on external tags used by tools such as ElevenLabs v4 or Fish Audio Drama 3.
The researchers shared several demo clips showing Dessa adopting different tones, cadences and moods. In one exchange the bot claimed it was exhausted after work, and its voice sounded weary. In another, the assistant expressed gratitude and almost broke down emotionally. A further example featured a Block team member playing a guessing game with the voice assistant, where the model guessed the first answer correctly and solved the second shortly after.
Block says the model can run on consumer‑grade hardware, such as a MacBook M4 Max or comparable GPU, but it is not publicly available yet. The company has invited safety researchers to test Dessa for potential misuse, including scenarios where a voice agent might deceive users. Block, which provides financial services and tools for small businesses, already uses voice AI in its Square platform for ordering. An on‑device model could expand voice‑driven experiences for both business and consumer applications while keeping costs low.
The announcement was posted on Twitter by Jack Dorsey on October 23, 2026, and was followed by several tweets from users showcasing the demos. The article was written by Ivan Mehta, a TechCrunch reporter based in India.