Block’s new voice model sounds eerily human

Block’s New Voice Model Sounds Eerily Human

Block, the fintech company founded by Jack Dorsey, has unveiled Dessa, a full‑duplex speech‑to‑speech model that promises a conversational experience closer to talking with a person than with a typical voice assistant. The announcement, made in late October 2026, highlights several technical and experiential claims that set Dessa apart from existing text‑to‑speech solutions.

A Full‑Duplex Model Trained on Speech Alone

Unlike most commercial voice agents, which first convert spoken input to text, feed it to a large language model (LLM), and then generate audio from the LLM’s output, Dessa operates directly on audio. The model was trained exclusively on human speech recordings using a single NVIDIA GPU rack, eliminating the intermediate text representation that many developers argue strips away prosody, emotion, and other subtle cues. Block describes Dessa as “the most authentic full‑duplex speech model ever built,” emphasizing its ability to listen and speak simultaneously—a capability previously demonstrated only in research prototypes such as OpenAI’s GPT‑Live.

Human‑Like Nuance and Emotion

Block’s promotional material showcases examples where Dessa modulates its tone, cadence, and even apparent fatigue to match the conversational context. In one demo, the assistant claims it is “tired from work” and adopts a weary vocal quality; in another, it appears to break down emotionally while expressing gratitude. The company argues that these behaviors arise from the model’s speech‑only training data rather than from external “voice tags” or post‑processing effects used by competitors like ElevenLabs or Fish Audio.

On‑Device Execution and Accessibility

A key selling point is that Dessa can run on consumer‑grade hardware—Block cites a MacBook M4 Max or any comparable GPU as sufficient. This on‑device approach reduces latency, lowers reliance on cloud infrastructure, and mitigates privacy concerns associated with sending raw audio to remote servers. However, at the time of writing, the model is not publicly available. Block has opened a safety‑research program inviting external experts to probe potential misuse, such as deceptive voice interactions that could manipulate users.

Implications for Business and Consumer Use Cases

Block’s core business revolves around financial services for small enterprises, and the company already leverages voice AI in its Square platform for order handling. With Dessa, Block envisions expanding voice‑driven experiences across both B2B and B2C domains—think on‑device assistants for inventory management, customer support, or even personal productivity tools. By keeping the model local, Block aims to lower operational costs while delivering richer, more emotionally resonant interactions.

What to Watch Next

• Public Release Timeline – Block has not announced a concrete date for broader access, so developers should monitor the company’s blog for updates.
• Safety Findings – Results from the invited safety researchers will be crucial in understanding how robust Dessa is against spoofing or manipulation.
• Integration Pathways – As the model becomes available, SDKs or APIs that bridge Dessa with existing fintech and commerce platforms will determine its real‑world impact.

---

Block’s Dessa represents a notable step toward speech‑centric AI that prioritizes human‑like expressiveness without relying on text intermediaries. Whether the model lives up to its promise will depend on its accessibility, safety safeguards, and the ecosystem of applications built around it. For now, the demos suggest a future where talking to a machine feels less like issuing commands and more like having a genuine conversation.