Block introduced the Dessa voice model on October 23, 2026, describing it as a full-duplex speech-to-speech system trained entirely on human speech. The financial technology company says Dessa can run locally on consumer-grade computers, with a MacBook M4 Max or better as the stated requirement. There is no public access to the model at present.
What full-duplex actually means
Most voice systems operate in turns: the user speaks, the system waits, then responds. Dessa is built to talk and listen at the same time. Block frames that simultaneity as the core technical distinction, and says talking with Dessa feels human-like. The company also claims it is the most authentic full-duplex speech model available.
The approach differs from the common pattern of pairing a speech interface with a large language model for knowledge. Dessa is speech-to-speech, not text-mediated, and Block researchers say its training data consisted entirely of speech. The company says whatever Dessa knows comes from its conversational training data.
Training on a single GPU rack
According to Block, the model was trained using one NVIDIA GPU rack. That detail sits alongside the local-execution claim as part of the company’s efficiency story: a model built on a modest hardware footprint, then deployed on hardware a developer or small business might already own.
One Block researcher described the work as a GPT3 moment for speech. The comparison is the researcher’s characterization, not a benchmark result. Block has not disclosed the model’s exact size or architecture, nor how its emotional behavior is steered.
Emotion without a text pipeline
Block argues that conventional voice AI throws away information that is not text — tone, hesitation, and the feelings behind a sentence. In the company’s account, a system that cannot hear a user’s emotional state produces worse interactions and potentially dangerous ones. Block calls the gap between powerful AI and machines that cannot perceive feelings ridiculous as well as risky.
Dessa is presented as an answer to that gap. Block says emotion is built into the model, and that it produces laughs, pauses, cries, and other emotional nuance. Whether those expressions are triggered by tags or by some other steering method has not been disclosed. Some competing systems, including ElevenLabs v4 and Fish Audio Drama 3, are known to use voice tags for emotion.
Researchers posted example conversations the same day as the announcement. In one, the bot sounded tired while discussing work fatigue. In another, it sounded emotional and nearly broke down. A third example was a guessing game with a Block team member named Cara; Dessa solved the first answer immediately and the second quickly.
Safety testing before any release
Block has not announced a release date or pricing. Instead, it issued a call for safety researchers to test Dessa, with attention to the risk that realistic voice agents could deceive callers. No results from that testing have been made public.
The caution is notable given how closely the model’s emotional realism could resemble a human on the other end of a call. Block’s own framing — that current voice interactions can be poor and that realistic systems raise deception concerns — places safety evaluation ahead of any commercial rollout.
Why a finance company is building voice AI
Block operates in finance and small-business solutions, and its Square payments platform already uses voice AI for order handling. An on-device model could lower costs for both business and consumer experiences by reducing reliance on cloud inference, though Block has not said whether Dessa will be integrated into Square products.
The broader market is moving quickly. OpenAI released its GPT-Live full-duplex model earlier in 2026, and voice AI has become a active area of competition. Block’s entry is distinguished less by scale than by its constraints: speech-only training, a single GPU rack, and execution on a laptop. Whether that combination holds up under independent scrutiny is what the safety testing period is meant to establish.
[…] Block Unveils Dessa, a Speech-Only Voice Model That Runs on a Laptop […]