Block, the company behind Square and Cash App, showed off a new voice model on October 23 called Dessa, and the demo clips spread fast because the voice on the other end laughs, pauses, sighs and talks over you the way a real person does. Jack Dorsey called it "the most authentic full-duplex speech model ever built," and he added that "talking with it feels like talking with a person." Below is what Block has actually shared about Dessa, what you can and can't do with it yet, and why people building with AI are paying attention.
What Dessa is
Dessa is a speech-to-speech model, which means audio goes in and audio comes out with no text step in the middle. Most voice assistants you've used work as a pipeline, where one model turns your speech into text, a language model writes a reply, and a third model reads that reply out loud. Block's official Dessa page says that approach loses a lot of what makes conversation human, so Dessa was built to hear and produce emotion, sarcasm, laughter and sighs directly from the sound itself.
The bigger claim is about training data. Block says it trained Dessa from scratch on nothing but human conversation, with no text and no pretrained models underneath, and that it learned language from what the company describes as "hundreds of times fewer words" than large language models need. Block hasn't published figures behind that comparison, so treat it as the company's claim for now.
The name has history. Block's research lab was founded by people from Dessa, the Toronto deep learning startup, whose staff joined the company after Square acquired it in 2020. The model now carries that same name, and Block lists Ragavan Thurairatnam as the Dessa lead.
What "full-duplex" means here
Full-duplex is the feature that makes the demos feel strange in a good way. Dessa listens and talks at the same time, so it can react while you're still speaking instead of waiting for you to finish a sentence and then responding. Block says the model learned from real conversation when to jump in and when to backchannel, which is the little "yeah" or "got it" people say to show they're following along.
If you've ever tried to have a back-and-forth with a voice assistant and ended up talking over each other, you know why this is the hard part. Turn-taking is something people do without thinking, and most voice products still handle it with a pause timer. A model that picks up those cues from audio could make voice apps feel a lot less like dictation.
Can you run it on a Mac
Block says inference runs locally on a MacBook Pro with an M4 Max chip or better, and that it runs faster than real time on that hardware. For training, Block says it used a single NVIDIA GB200 NVL72 system, which is a rack-scale machine rather than one graphics card. Local inference on a laptop is a big part of why developers are interested, since it points toward voice features that work without sending audio to a server.
We looked at what it takes to run a large open model on Apple hardware in our guide to running Muse Spark Open on a Mac, and the same questions about memory and chip generation will apply once Dessa is in more hands. If the difference between downloadable models and hosted ones is still fuzzy, our explainer on open-weight vs closed AI models covers the basics.
How to try Dessa right now
You can't download or use Dessa yet. Block calls it a research preview, and the Dessa page says launch links are "coming soon" without giving a date. For now, early access is going to a small group of AI safety researchers who are testing for risks the team hasn't anticipated, and researchers who want in can contact Block through its page.
What you can do today is listen. Block has posted recorded samples on its site and says more recordings are coming on X. Keep in mind that every clip so far is a company demo, and Block hasn't shared the model's size, benchmark results or independent test results, so the real-world experience could differ from what you hear in a polished sample.
The scam concern
Not everyone heard the demos and felt excited. As CoinDesk reported, filmmaker Justine Bateman questioned why AI voices need to sound human at all, and journalist Shiv Aroor warned that scam calls could get much worse. The worry is easy to follow, because a voice that sighs and laughs on cue is harder to recognize as a machine over the phone.
Block's answer so far is the staged release. On its page the company says "a model that learned to talk from sound alone is new territory, and we want outside eyes on it," which is why researchers are getting access before the public. Block also argues that a model that hears tone directly could get better at noticing things like distress that a text transcript would miss. Whether those safeguards hold up is something to watch once launch details arrive.
Why this matters for group chats and voice apps
Voice is creeping into the apps people already use with friends, from voice notes to AI features that summarize long threads. We covered how chat apps are adding these tools in our look at AI in group chats, and a model like Dessa hints at the next step, where an assistant in a voice room could follow the rhythm of a conversation instead of waiting for silence.
It also feels like one more moment where real tech catches up to the movies. Plenty of the sci-fi movies that predicted today's AI imagined a voice you could talk to like a friend, and the reaction to Dessa shows people are both drawn to that idea and a little uneasy about it.
If you want to trade takes on the Dessa samples, or compare notes once Block opens access, the AI Trends Tribe is where builders and curious listeners are already posting their reactions. It's a good place to ask what people think the safety testing should cover, and to hear from folks who will actually try running it on their own Macs. For in-person versions of these conversations, follow the Tribe Events calendar on Luma, where we post AI talks and meetups around the Bay Area.
