Block’s newly introduced voice model sounds uncannily human-like

Jack Dorsey’s Block has released Dessa, a voice model from its research lab that runs locally and sounds strangely human through expressions such as laughs, pauses, and cries. The model was trained fully on human speech using a single NVIDIA GPU, and its full-duplex design lets it talk and listen at the same time. That matters because Block is pitching a speech model that feels more like a person than a text-driven voice assistant.

Background: Voice AI has been built around text

Block says current models can sound unnatural and annoying because text sits in the pipeline that drives speech generation. It points to systems such as ElevenLabs v4 and Fish Audio Drama 3, which use voice tags to drive emotion in speech generation. Dorsey’s company claims that emotion is built into Dessa, though it is not clear whether some kind of tags still steers that emotion.

The broader pattern is that speech models are typically paired with a large language model, or LLM, for knowledge retrieval. Block said Dessa’s knowledge stems from the conversational data it has. The comparison it makes is with OpenAI’s GPT-Live model, released earlier this year, which also represents a full-duplex approach.

Development: Block frames Dessa as full-duplex speech-to-speech

On October 23, 2026, Jack said Block created Dessa because voice is the most natural way to interact with intelligence. He called it the most authentic full-duplex speech model ever built and said talking with it feels like talking with a person, full of emotion, nuance, pauses, and a natural cadence.

Block’s researchers posted multiple examples of conversations with Dessa where the agent had different voices, cadences, and emotions. In one example, the bot said it was tired from work, and at that point it sounded tired. In another, the bot was talking about the person making it comfortable while sounding emotional and almost breaking down.

A separate post showed Block team member Cara playing a guessing game with the voice assistant. Block said the model guessed the first one right away, and the second did not take long either. Ragavan Thurairatnam described the result as a GPT3 moment for speech, saying Block made the first AI model you can have a realistic conversation with that was trained entirely on speech, and that Dessa is a full-duplex speech-to-speech model that can run on consumer GPUs. Aydin Polat also posted a test described as “Dessa=full-duplex speech model vs Human=me.”

In a blog post, Block said that interacting with current voice AI can be a bad experience because it turns voice into text and throws away information that text alone cannot carry. It said the most powerful AIs cannot hear how a person feels, and called that ridiculous, potentially dangerous, and a worse experience.

Latest progress: Local access is limited while safety testing begins

Block said the model can run locally on a consumer-grade computer such as a MacBook M4 Max or better. There is no public access to Dessa at the moment, and Block has put out a call for safety researchers to test the model. This is likely to test cases where a voice agent could trick people on the other end.

The business relevance is tied to Block’s core work in finance and selling solutions to small businesses. Square already uses voice AI to handle ordering. With its own voice model that can run on-device, Block says it can power a lot more business and consumer experiences at a lower cost.

The current stage is a private research release: demonstrations are public, access is not, and safety testing is the next gate before broader use. The single NVIDIA GPU training detail and the MacBook M4 Max or better requirement are the numbers that define the hardware boundary of the release.

Meta description: Block says its new Dessa voice model runs locally on consumer GPUs, uses a single NVIDIA GPU for training, and can speak and listen at once.

Tags: Block, Dessa, voice AI, full-duplex, NVIDIA GPU

Featured image: Abstract waveform over a dark circuit-like grid, no people, no company logos, no text.