AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Block’s new voice model sounds eerily human

Collected Oct 24, 2026

Block, the financial technology company led by Jack Dorsey, released a new voice model called Dessa from its research lab on October 23, 2026. According to the company, the model runs locally and produces speech with human-like expressions such as laughs, pauses and cries. Block says Dessa was trained fully on human speech using a single Nvidia GPU rack, and that it is a full-duplex model, meaning it can talk and listen at the same time.

Full duplex is the key architectural claim. Most voice assistants work in turns: the user speaks, the system transcribes the audio to text, a language model composes a reply, and a text-to-speech engine renders it back into audio. That pipeline introduces latency and, more importantly, discards paralinguistic information such as tone, pacing and affect. Dessa is described as speech-to-speech, skipping the text stage entirely. Block argues that current models sound unnatural because text sits in the pipeline driving speech generation, and that emotion in Dessa is built into the model rather than layered on top.

The comparison the company draws is with systems such as ElevenLabs v4 and Fish Audio Drama 3, which use voice tags to steer emotion during generation. Block claims its own approach bakes emotional expression in natively. The reported material leaves open whether any tag-like steering is still involved in Dessa, so the precise mechanism behind the model's emotional range is not settled.

Knowledge handling is another departure. Speech models are commonly paired with a large language model for retrieval and reasoning. Block says Dessa's knowledge comes from the conversational data it was trained on. In practice, that suggests a model optimized for the texture of conversation rather than for breadth of factual recall, and it implies a narrower scope than an assistant backed by a general-purpose LLM. A likely trade-off is that Dessa may excel at sounding present and responsive while being weaker at answering questions outside what its conversational training covered.

Block's researchers published several conversation samples to demonstrate range. In one, the agent said it was tired from work and sounded tired while saying it. In another, the bot talked about a person making it comfortable while sounding emotional and close to breaking down. A third shows a Block team member playing a guessing game with the assistant, which identified the first answer immediately and the second shortly after. The examples show different voices, cadences and emotions across conversations, which is the kind of variety that makes a voice model feel less like a fixed persona and more like a responsive interlocutor.

Hardware requirements are unusually modest for a model of this ambition. Block says Dessa can run locally on a consumer-grade computer, naming a MacBook M4 Max or better as the threshold. That is a meaningful change from cloud-hosted speech stacks, where every utterance carries inference cost and network latency, and where audio leaves the device. On-device operation addresses both: cost per interaction drops toward zero at the margin, and audio need not be transmitted.

Access is limited. There is no public availability at the moment, and Block has issued a call for safety researchers to test the model. The most obvious area for that testing is deception and impersonation: a voice agent convincing enough to pass as human on the other end of a call creates fraud and social-engineering risks, which is squarely relevant for a company whose core business is payments and small-business finance. Testing for those cases before wider release is a sensible sequencing, and it is an inference rather than a stated company rationale that this is the purpose of the call.

The strategic logic is easy to see. Square already uses voice AI to handle ordering, and an on-device model would let Block power more business and consumer experiences at lower cost than paying per-token or per-minute rates to a third-party provider. Owning the model also means owning the voice, the latency profile and the data path.

Why it matters: developers building voice interfaces have largely depended on cloud APIs that chain transcription, an LLM and speech synthesis, with the associated latency and per-call costs. A locally runnable, full-duplex speech-to-speech model from a major company signals that this stack may compress into a single on-device component, which would change cost models and privacy postures for anyone shipping voice features. The absence of public access means that shift is not yet actionable; the safety review and whatever release follows will determine whether teams can plan around it. Given Block's payments footprint, expect scrutiny of how convincingly Dessa can imitate a person, since that capability cuts both toward better experiences and toward harder-to-detect fraud.

Read at TechCrunch · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Typically, speech models are paired with a large language model for knowledge retrieval. Block said that Dessa's knowledge stems from the conversational data it has.