The Next Step for Voice AI Is to Throw Out Text—at the Cost of an M4 Max

Block's Dessa Lab has released a full-duplex voice model called Dessa. Audio in, audio out, without going through speech-to-text, and it can run locally on a MacBook Pro M4 Max or better.

Chubby's take in the tweet was blunt: this voice model can run locally on an M4 Max Mac or better, and it feels like magic. In the demo, two waveforms—one for the human and one for Dessa—sit side by side, with buttons labeled laugh, surprise, and speak louder. Dessa's laugh happens mid-sentence, like a person who has heard something funny; ask it to speak louder, and it raises its volume. The official site says no one taught it these things.

Jack Dorsey's line is that they believe voice is the most natural way to interact with intelligence, so they built Dessa, billed as the most realistic full-duplex voice model, with emotion, detail, pauses, and natural rhythm. The official site puts the problem with existing voice AI more bluntly: converting sound to text first loses a huge amount of information that text cannot carry. Dessa was trained from scratch using only the sound of human conversation, with no text and no pretrained model. It listens and speaks at the same time, can interject, and can use backchannels like 'yeah,' 'got it,' and 'ooh' to show it is listening. Inference is faster than real time, so there is no awkward waiting.

Dessa's way of learning is also worth a look. The official site says it learns hundreds of times fewer words than an LLM—of course, LLMs also have to learn C++ and Angular; that's the site's self-deprecating joke. Speech is not repeatable: say the same sentence twice, and the data is different each time; a spoken word has thousands of times more bytes than the same word in text. Dessa also has to learn accents, pauses, breathing, sarcasm, humor, intonation, emotion, and when to interject. The official site says training used just 1 NVIDIA GB200 NVL72. Don't be fooled by 'one'—that's a rack-scale system.

The official site also includes a recording of life advice, and Dessa sounds empathetic. All demos are real-time recordings of a person talking to Dessa, with no one operating it; captions were added afterward. It can choose what it says, when it says it, and how it says it—even its voice and personality are its own.

Can you use it now? Not freely. The official site has opened a waitlist and is also giving early access to a small number of AI safety researchers; the application form requires an institutional email and a Google Scholar or blog link. Some commenters pointed out that there is currently no public trial—only recordings and a few safety researchers. This limitation matters. It is more like a technology demo and an entry point for safety research than a consumer product that can replace Siri tomorrow.

The local-running angle blew up the comments section the most. M4 Max or better immediately rules out most laptops. One commenter said this amounts to buying a $4,000 Mac first, after which voice costs zero per minute. Others worried about the fans, saying the Mac started sweating when it heard this demo. Someone else added: Siri still needs WiFi to set a timer. That is also where the value of a local model lies: once it runs on a user's machine, there is no server that someone can just switch off.

The comments quickly slid into identity verification. One commenter brought up the Lindy rule: humans only talk to living things, including other people, themselves, animals, and objects with spiritual significance. Now that computers can chat too, this rule needs a patch. Others asked whether phones need blockchain authentication, otherwise you can't tell whether the person on the other end is human. Some said this is a new Black Mirror episode, and some are already waiting for an AI girlfriend.

Based on the information made public, what makes Dessa worth watching is its training approach: throw out the text intermediary and learn directly from human conversation, preserving tone, pauses, interruptions, and emotion. This path is harder and closer to what voice interaction is supposed to be. But at this stage it is still out of reach for ordinary users: the barrier is a high-end Mac, and the entry point is a waitlist and a safety research application. You can treat it as magic, but don't delete Siri just yet.