Block Dessa Voice AI Learned to Talk From Speech Alone and Runs on a MacBook

In one of the demo clips Block published this week, a voice AI laughs halfway through its own sentence, then picks the thought back up as if nothing happened. That model is Block Dessa, and the company behind Square and Cash App says it learned to talk without ever reading a single word of text.

Block, the payments company run by Jack Dorsey, unveiled Dessa on October 23 as a research preview. According to Block's announcement and reporting from CoinDesk, Crypto Briefing and Notebookcheck, the model was trained from scratch on recordings of human conversation only. There was no text data and no pretrained language model underneath it. Block also says it runs locally on a MacBook Pro with an M4 Max chip or better, with no cloud server needed to answer you.

Dessa skips the text step entirely

Most voice assistants today work as a relay. Your speech gets turned into text, a language model writes a reply, and a second system reads that reply aloud. Dessa is built as a speech to speech model, so it takes in audio and produces audio directly.

Block describes it as full duplex, meaning it can listen and speak at the same time. In practice that lets it interrupt, drop in a quick acknowledgment while you are still talking, back off when you cut in, and react before you finish a sentence. Block claims it also picks up cues that a transcript would lose, such as sarcasm, hesitation, laughter and changes in emotional tone.

The company's pitch for training on audio alone is simple. "Babies learn to talk without ever seeing a word written down," Block wrote, adding that a model can learn the same way, though "it's just very hard." Crypto Briefing reports that Block says Dessa learned language from far fewer words than text based models need, but the company did not publish a detailed comparison.

One Nvidia rack for training, one laptop for running

The hardware claims are the part that has drawn the most attention from developers. Block says Dessa was trained on a single Nvidia GB200 NVL72 system, a rack that holds 72 Blackwell GPUs. That is a large machine, but it is modest next to the clusters used for frontier text models.

On the other end, Block says inference runs faster than real time on an M4 Max MacBook Pro, with no noticeable pauses in conversation. If that holds up, it means a humanlike voice agent could live entirely on a personal computer. Several commentators on X pointed out the flip side: once a model like this runs offline, no server sits in the middle to monitor or shut it down.

Block has not disclosed the model's parameter count, released its weights, or shared standard benchmark scores. Every demo so far was selected and published by Block itself, and none has been independently tested.

What the demos show, including the misses

The clips posted by research lead Ragavan Thurairatnam cover chats about movies, reactions to personal stories and guessing games. In one, Dessa finishes a sentence for a speaker who is struggling to remember a film title. In another, it talks with OpenAI's GPT Live 1 voice model. Block says the conversations were recorded live with no one steering the responses.

Crypto Briefing notes the samples are not flawless. Some show Dessa misidentifying who is speaking or changing the subject without warning.

The team behind the model is small. According to CoinDesk and Crypto Briefing, it reports directly to Dorsey and includes people from Dessa, the Toronto AI startup Square acquired in 2020, which is where the model's name comes from. Dorsey wrote on X that talking with it "feels like talking with a person."

Safety researchers get it first, and scam worries arrive early

Dessa is not available to the public. Block says it is inviting a limited group of AI safety researchers to evaluate the model, and everyone else can join a waitlist. No release date has been given.

The reaction split quickly. Binance founder Changpeng Zhao suggested on X that a voice this natural could help with teaching children. Others were less enthusiastic. Filmmaker Justine Bateman asked why AI voices need to sound human at all, and journalist Shiv Aroor predicted scam calls would reach "pandemic level," as CoinDesk reported.

Those concerns are not abstract. Voice cloning is already a common tool in family emergency scams, and our guide to spotting AI deepfake scams aimed at your family covers the habits that still work, like a family code word. A model that can laugh, hesitate and interrupt on cue would make those calls harder to catch.

Block frames the emotional side differently. The Dessa team argues that AI "is still nearly deaf to the emotion in our voices," and says a speech native model could eventually get better at noticing distress or pain that a text transcript misses. The company says the outside researchers are there partly to find risks it may have overlooked.

The release lands in a busy week for AI safety moves. OpenAI just said it will train frontier models in an offline copy of the web, and Anthropic recently warned about AI worms spreading through smart devices. Voice is also where the big assistants are heading, from ChatGPT's GPT 6 interface changes to more agent style tools.

The open question is whether Dessa's demos survive contact with independent testers. Until Block shares benchmarks or weights, the laptop claim and the humanlike timing rest on the company's word. More AI coverage is on our AI news page.