Block has introduced Dessa, a conversational artificial-intelligence model the company says was trained from scratch using human speech rather than text or a pretrained language model. The research preview accepts audio and produces audio directly, allowing it to listen and speak at the same time. Block's release page includes recordings in which the model interrupts, laughs, changes volume and reacts to vocal emotion, but the demonstrations were produced by the company and do not constitute independent performance tests.

A model built around speech rather than transcripts

Most voice assistants connect several stages: speech recognition converts a person's words into text, a language model generates a response, and a speech system reads that response aloud. Block says Dessa removes the text stage. Everything the model knows came from recorded conversation, including timing, pauses, accents and intonation. The model is full-duplex, meaning it can receive and generate speech concurrently instead of waiting for a speaker to finish every turn. Block also says it trained Dessa on one Nvidia GB200 NVL72 system and that inference runs faster than real time on a MacBook Pro equipped with an M4 Max chip or better. Those hardware and speed statements are company claims, not independently published benchmarks.

Related reporting: Revolut AI Strategy: PRAGMA, Research and Plans Beyond Banking

The release is a research preview, not a product launch

Block has not released model weights, a public API, pricing or standardized evaluation results. The company says it will give a small group of AI safety researchers early access and asks applicants to provide an institutional email, with a research profile or blog optional. TechCrunch and The Verge independently confirmed that access is limited and that the public material consists mainly of first-party demonstrations. That distinction matters: the recordings show what Block chose to present, but outside researchers have not yet established how reliably Dessa handles noise, unfamiliar accents, factual questions, adversarial prompts or long conversations.

Why a payments company is working on voice AI

Block operates Square, Cash App and other financial-technology businesses, and Square already offers voice-based ordering tools for restaurants. A compact model that can run locally could eventually reduce cloud dependence or improve the responsiveness of customer-facing systems. However, Block has not announced a Dessa integration with Square, Cash App or any financial product. Any commercial use therefore remains a possibility rather than a disclosed rollout. The immediate development is the creation of a separate research model inside Block, not a change to customer accounts, payments or merchant services.

Natural speech raises both interface and safety questions

Block argues that direct audio processing preserves information that text transcripts discard, including hesitation, sarcasm and distress. That could make voice systems more responsive to how a person speaks, not just the words used. Greater naturalness also creates risks: listeners may find it harder to recognize that they are speaking with software, and convincing synthetic speech can be misused for deception. Block says external safety researchers are being invited to look for risks the team has not anticipated. The useful next evidence will be independent evaluations, clear disclosure practices, reproducible latency and quality measurements, and details about access controls. Researchers will also need to test whether the model preserves its apparent responsiveness outside Block's curated examples. Until those arrive, Dessa should be understood as a technically notable but unverified preview rather than a broadly available assistant.

Sources

AI-generated editorial image; not a photograph of the reported event. Prepared with AI assistance and source verification.