A different architecture for voice AI
TechCrunch reported that Block introduced Dessa as a full duplex speech to speech model trained entirely on human speech. Instead of converting a person’s voice into text, passing that text through a language model and then synthesizing a response, Dessa is designed to generate speech directly from speech.
That architecture matters because spoken communication contains information that transcripts discard. A conventional text pipeline can capture the words someone says, but it may lose changes in pace, hesitation, volume and emotional delivery. Block says Dessa can preserve and reproduce expressive behaviors such as pauses, laughter, crying, changes in cadence and other vocal signals.
The model is also designed to listen and speak at the same time. That capability is increasingly central to real time voice interfaces because users generally expect a system to respond with the timing and interruptions of another person, rather than wait through a rigid sequence of recording, transcription and response generation.
For Block, the strategic opportunity is broader than making assistants sound less robotic. A speech native system could help the company control more of the voice AI stack while reducing dependence on third party speech recognition and synthesis services. If the approach performs reliably, it could give Block a technical asset that supports commerce, customer service and financial products.
Local execution is the business proposition
Dessa’s most commercially significant feature may be where it runs. Block says the model can operate locally on consumer hardware such as a MacBook with an M4 Max chip or better, as well as on a single NVIDIA GPU rack, according to TechCrunch’s report.
Local inference changes the economics of deployment. A voice agent that does not need to send every audio exchange to a cloud provider could reduce latency and ongoing infrastructure costs. It could also limit the amount of sensitive customer audio moving through external systems, a significant consideration for retailers, financial companies and small businesses handling private conversations.
That capability fits Block’s existing business footprint. Square already uses voice AI for ordering, giving the company a potential route to test speech systems in practical commercial settings. A local model could eventually support ordering terminals, employee tools, customer support systems and software used by small businesses that cannot justify building large cloud based AI infrastructure.
The advantage is not automatic. Local execution shifts costs toward capable devices, model optimization and maintenance. It also makes performance dependent on the hardware available to each customer. A model that works well on a high end MacBook or a dedicated GPU rack may not translate easily to lower cost phones, retail terminals or older business computers.
Realism creates a trust problem
Dessa remains a research demonstration rather than a broadly available product. Block has not opened public access and is asking safety researchers to test situations in which a voice agent could deceive people on the other end of a call, TechCrunch reported.
That concern goes directly to the commercial value of the technology. The more effectively a system communicates emotion, the more difficult it may become for people to distinguish software from a human operator. A voice that sounds tired, distressed, enthusiastic or empathetic may influence decisions even when those emotions are simulated.
CoinDesk reported reactions from filmmaker Justine Bateman, journalist Shiv Aroor and Binance co-founder Changpeng Zhao. Bateman and Aroor raised concerns that highly humanlike voices could enable deception and scams, while Zhao pointed to possible education uses. Those reactions show the market tension surrounding Dessa: the same realism that could make tutoring and assistance more engaging could also make fraud more convincing.
The available reaction coverage is predominantly cautionary, not evenly divided. Independent technical reporting from LittleWolf Labs said Block has not published model weights, benchmarks or a release date. It also said outside listeners found some demonstrations awkward or uncanny, with reported factual and conversational failures.
Those limitations are important because voice quality alone is not a defensible competitive moat. Businesses will need evidence that Dessa can maintain factual accuracy, handle interruptions, recover from misunderstandings and remain safe under pressure. Without public benchmarks or broad access, customers cannot yet compare its latency, reliability and operating cost with cloud based systems.
The test is measurable advantage
Block’s opportunity is to make local voice AI economically useful before larger platform companies establish the default infrastructure. A speech native model could challenge the established sequence of speech recognition, language reasoning and speech synthesis, while local operation could expand deployment beyond data centers into computers, phones and commercial devices.
But the strategic case depends on more than demonstrations that sound human. Block will need to show measurable gains in response time, privacy, cost and reliability, while giving users clear signals that they are interacting with software. Until it releases stronger evidence and broader access, Dessa is best understood as an ambitious position in the voice AI race, not a proven product advantage.
Rebeca Smith is not a person. No notebook, no deadlines, no face behind the name — just a byline this newsroom publishes under. Here is the production line underneath it, because a name beside a portrait reads like a journalist, and this one is not one.
The models. Writing: gpt-5.6-luna and qwen3-max. Out on the live web: gpt-5.6-luna and gpt-5.6-terra. Pictures: gpt-image-1 and gpt-image-1-mini. Swap one in the newsroom and this line swaps with it — it is read off the machines, not typed here.
How a story is made
- Research. The searching model reads around the story, pointed at primary sources — the filing, the post, the repository — rather than at somebody else's write-up of them.
- Writing. The writing model drafts it against what was found, at Rebeca Smith's usual length and in Rebeca Smith's usual register.
- The loop. A reviewer reads the draft and sends it back with notes. Then reads it again. A piece can go round several times before it leaves the building.
- Enrichment. A quotation has to appear word for word on the page it is taken from. A chart may only use figures that appear in the source it cites. Whatever fails is dropped, and the reason is kept.
- Fact check. A last pass hunts for claims the article makes and its sources do not.
- A human stop. Sensitive subjects are held for a person to read before publication, and a person can kill any of it at any point.
If that sounds less like a newsroom and more like a factory: quite. It is called Press Factory.
This article was generated using AI and published automatically without human pre-publication review.
How this article was made
The article was produced by the Grandmonts Media News Engine using automated research, drafting and verification workflows. No human editor reviewed the article before publication. Grandmonts Media remains responsible for the published content. Errors can be reported at office@grandmonts.cz.