Block's Dessa sounds human. Here's what it still needs before it answers a shop's phone

By Ross Michaels, azlabs

On October 23, a small research lab inside Block (the company behind Square and Cash App) showed off Dessa, a voice model that learned to talk only by listening to people. If you heard the clips, you know why it took over X that afternoon. Jack Dorsey's post about it had passed 12.8 million views by Sunday morning.

We build AI voice agents that answer the phone for home service shops, med spas, and injury firms, so we listened to every clip on Block's page. Here's what Dessa is, what it isn't yet, and what it would need before it could pick up a real business line.

What Dessa is

Block calls Dessa "a true speech-to-speech model." A lot of voice AI still turns what you say into text, writes a reply, and reads that reply back to you. Block says Dessa skips text completely, so the tone of your voice, a sigh, or a laugh goes straight in and comes straight back out.

It's also full duplex, which means it listens and talks at the same time. In Block's words, it "knows when to cut in and when to backchannel ('yeah!', 'got it!', 'ooh')" the way a person on the other end of the line does.

The training is the unusual part. Block says it trained Dessa "from scratch on nothing but human conversation," with "no text, no pretrained models," and that it learned language from "hundreds of times fewer words" than a text model needs. The whole thing was trained on one NVIDIA GB200 NVL72, a single rack of GPUs, and Block says it runs locally on "a MacBook Pro M4 Max or better" faster than real time. (Block)

Ragavan Thurairatnam, who leads the Dessa team, called it "a GPT3 moment for speech."

What it isn't yet

As of October 25, Dessa is a research preview. There's no API, no downloadable model, no license, and no price. The Download button on Block's page says "Download coming soon," and the waitlist says it "opens soon." The only people getting in are a small group of AI safety researchers, who apply with a university, company, or lab email. (Block, TechCrunch)

Block also hasn't published benchmarks, a latency number, a model size, a list of languages, or a paper. What we have is 21 recordings that Block picked. So nobody can put Dessa on a phone line today, us included.

What the recordings show

A lot of it is genuinely impressive. Dessa laughs mid-sentence when something is funny, gets louder when you ask it to speak up, and finishes your sentence when you blank on a movie title. In one clip it talks with OpenAI's GPT-Live-1, and The Verge thought Dessa came across as the more natural of the two.

It also gets things wrong in very human ways. In another, it opens by calling the person Ronaldo, which isn't his name, as The Verge pointed out.

The clip we keep coming back to is called "Refund." A caller says a bag she ordered arrived damaged and asks whether she can return it for a refund. Dessa, playing the store, says yes, then tells her she can get half her money back, adds "I know the return policy makes no sense," and finally agrees to swap the bag with "no extra charges" when she pushes back. It sounds completely natural, and it made up a return policy on the spot.

That isn't a knock on a research preview. Block says nobody steers these clips, and Dessa had no store, no policy, and no order system behind it. It did exactly what it was built to do, which is hold a natural conversation. It's just the cleanest example we've seen of the gap between sounding right and being right.

Sounding human was never the hard part of a business call

Someone calling about a dead AC unit doesn't need the agent to laugh at the right moment. They need it to get the address right, find a real open slot, and actually put the job on the schedule. Two tests that came out the same week as Dessa show how hard that still is for voice models you can already buy:

• Sierra's τ-Entity asks a voice agent to capture one detail per call, like a name, phone number, or address. The best model, GPT Live 1, got it right on 60.3% of regular calls, and only 34.0% of tasks passed in all three conditions (regular, noisy, and lots of background talking). (Sierra)
• AudioAgentBench, from DesignArena, ran seven speech-to-speech models through 420 runs of customer tasks. The best overall pass rate was 80.3%, but on turns where the agent had to actually do something (call a tool), the best score was 61.4%. A scheduling task with similar-sounding names and details that changed mid-call averaged 32.3%. (DesignArena)

Those are models with public APIs and tool calling. Dessa hasn't been through either test, and Block hasn't said whether it can call tools at all.

What Dessa would need before it answers a shop's phone

Here's the bar we think any voice model should clear before it talks to a shop's customers, Dessa included:

1. A way in. An API or a model we can run, with a license that allows business use.
2. The shop's rules. A way to give it the service area, hours, prices, and policies, so "half your money back" never comes out of its mouth.
3. Real actions. It has to check the calendar, book the job or hand it to your team, text a confirmation, and get emergencies to a person, and it can never say a job is booked when it isn't.
4. The details right. It needs to read back names, numbers, and addresses, and be measured on that in a noisy room, not only in a quiet one.
5. Real phone calls. Tested on actual phone lines, with a bad connection and a dog barking, not just a laptop mic.
6. Honesty. It tells callers it's an AI. Block's own safety call-out exists because a voice this human raises real questions, and TechCrunch points to the risk of a voice agent tricking people.

Why we're still excited

The full-duplex part is real progress. Talking over each other, the "uh-huh" while someone explains, hearing that a caller is frustrated: those are the moments where callers give up on today's voice agents. A model that handles them natively would make every call feel shorter.

If Block opens Dessa up to developers, we'll put it through the same kind of calls we care about (an address in a loud kitchen, a booking that has to land) and share what we find.

If you're picking a voice agent today

Skip the question of how human it sounds. Ask for a recording of the agent taking an address over background noise, and ask to see the booking it says it made. Those two things will tell you more than any demo.

If you want to hear how Front Desk handles that kind of call, you can try a live demo at front-desk.ai.