Block unveils Dessa, voice AI trained only on speech

By VK

Published 25 Oct 2026, 11:43 UTC

Reporting from The Verge, TechCrunch

Block's research lab built Dessa, a full-duplex voice model it says was trained without text and runs on a high-end MacBook. Safety testers go first.

What it means for founders

A simpler voice stack may be coming. Many voice products today chain transcription, a language model and text to speech from separate vendors such as ElevenLabs, paying for each step. One local model doing all three could cut latency and per minute costs, so long contracts for any single layer deserve caution.
On-device voice changes the privacy pitch. A model that runs on a laptop could keep audio off the cloud, which matters for health, finance and support calls. Expect buyers to ask whether yours can run locally.
Human-sounding voices raise the disclosure bar. Callers may struggle to tell a lifelike agent from a person. Plain AI disclosure looks like a sensible default, and call recording consent rules vary by place, so they are worth checking with a lawyer.
Watch the launch links. The signal is whether Dessa ships as open weights, a paid API or a Block-only feature. Rivals are moving: ChatGPT Voice now acts on email and calendars.

The story

Jack Dorsey's Block has unveiled Dessa, a conversational voice model that learned everything it knows from recordings of people talking. Block announced it on Friday, October 23, through a project page and on X. Nobody outside Block can use it yet; a handful of outside safety researchers will get early access, the company says.

How Dessa differs from other voice AI

Most voice assistants turn speech into text, have a language model reply, then convert the reply back to audio. Block argues that pipeline discards tone, timing and feeling. Dessa removes text from the loop: sound goes in, sound comes out, and the company says it was trained from scratch with no text data and no pretrained model underneath.

It is also full duplex, so it listens while it speaks. Block says that lets it interrupt, laugh partway through a sentence or slip in small acknowledgements, as a human listener would. TechCrunch compared that ability with OpenAI's GPT-Live, which also talks and listens at once. Block says Dessa produces speech faster than real time and needs no cloud server: an M4 Max MacBook Pro or a more powerful computer can host it. Training used a single NVIDIA GB200 NVL72 rack, according to the company, which also claims Dessa picked up language from a word count hundreds of times smaller than large language models require.

Dessa lead Ragavan Thurairatnam called it "a GPT3 moment for speech" on X, as TechCrunch reported.

What the demos show

Block's clips, which it says are live and unsteered, include a guessing game, a chat about Christopher Nolan films, an emotional reaction to a surprise and a conversation with OpenAI's GPT-Live-1. The Verge found some exchanges uncanny, including one where the model opens by calling the speaker by the wrong name.

The lab behind Dessa reports to Dorsey. Block calls it a very small team whose founders came from Dessa, a deep learning startup in Toronto. It lists curiosity, closer understanding between humans and machines, and safety as its reasons, arguing that an assistant unable to hear pain in a voice could carry on regardless.

What we don't know yet

When developers can use Dessa, and under what licence or price. The project page says only that launch links are coming soon.
Whether its emotional delivery is shaped by hidden tags or controls, a question TechCrunch raised.
How it scores on accuracy or reasoning. Block has shared demos, not benchmark figures or a model card.
Whether Block will put it inside Square, which already offers AI voice ordering.