Tech brief

Block's New AI Voice Sounds Startlingly Human. What Lifelike AI Voices Mean for Phone Scams

By , Tech Cheat Codes Team

Published Updated 6 min read

Dark text card with white lettering reading: This AI voice sounds human
On this page · 3 sections
  1. Why it sounds so different
  2. The scam problem
  3. References

In one of the clips Block posted on Friday, October 23, someone surprises an AI voice called Dessa with an anniversary trip to Paris. It gasps "oh my god" under its breath, then answers with what Block describes as "teary gratitude." In other clips it laughs in the middle of a sentence and jumps in with "yeah!" and "got it!" while the other person is still talking.

Dessa comes from a small AI research lab at Block, the company behind Square and Cash App, co-founded by Jack Dorsey. The lab built it only from recordings of people talking. You can't download it or call it. It isn't available to the public, and for now only a small group of outside safety researchers is getting early access, so nothing changes on your phone today.

Block describes every clip as a real, unscripted conversation, with subtitles added afterward. When a person blanks on a movie title, Dessa finishes the sentence. When someone asks it to speak up, it gets louder. And it laughs the moment something is funny, mid-sentence, instead of waiting for the other person to finish. One clip is a conversation between Dessa and GPT-Live-1, OpenAI's voice AI, so both voices are made by AI. (Block notes that OpenAI isn't affiliated with Block and hasn't reviewed or endorsed Dessa.)

Not every clip is smooth. Some of the chats come across as a little awkward, and in one, Dessa starts out by mistaking the person it's talking to for someone named "Ronaldo." Jack Dorsey shared the release on X, Block's researchers posted more clips there, and tech sites including The Verge and TechCrunch covered it.

Why it sounds so different

Most voice assistants work in three steps. They turn what you say into text, work out a text answer, then read that answer out loud in a computer voice. Anything that doesn't fit in text gets lost along the way, whether it's a laugh, a sigh, sarcasm or a shaky voice. That's a big reason most AI voices still sound flat or slightly off.

Dessa skips the text step entirely. Sound goes in and sound comes out. Block trained it from scratch on nothing but recordings of human conversation, with no written text and no existing AI model as a starting point, which means everything it knows, including the facts it brings up, came from listening.

It can also listen and talk at the same time, the way people do on a real phone call. That's what lets it cut in and react while you're speaking, or drop in a little "mm-hm" to show it's following along. Older voice assistants mostly wait for you to stop before they answer.

It runs right on the laptop itself, as long as that laptop is a MacBook Pro with an M4 Max chip or better. Block trained it on one NVIDIA GB200 NVL72, a single rack of AI computer chips. What's missing is a price and a release date, since Block's announcement lists neither. Its page shows Download and waitlist buttons, but those say they're coming soon. The announcement also doesn't say how big the model is or how it scores on standard tests, so until outside experts test it, the only public evidence is Block's own demos.

Block gives three reasons for building it: curiosity, helping people and AI understand each other better, and safety. The safety argument is that today's AI is "nearly deaf to the emotion in our voices," and a system that can actually hear distress or pain in someone's voice could respond better to it. Block is also openly asking for outside help and giving early access to a small group of AI safety researchers "to find the risks we haven't anticipated." For a voice this convincing, that's a sensible step.

There's a business angle, too. Block runs Square, which sells its products to small businesses, plus Cash App, and Square already uses voice AI to handle ordering. If Block eventually puts a voice like Dessa into its products, customers calling a business could end up talking to an AI that sounds a lot like a person. Block's announcement doesn't mention any such plan, so for now that's only a possibility.

The scam problem

Block hasn't released Dessa, and nothing suggests it's being used for scams. But AI voice tricks aren't new. The Federal Trade Commission (FTC), the U.S. agency that handles scam complaints, has warned since at least 2023 that scammers can copy a loved one's voice from a short clip posted online and use it in fake "family emergency" calls. A caller who sounds just like your grandson claims he's in trouble and asks for money fast.

Once AI voices come with real laughter and pauses and emotion, "it sounded just like them" stops proving much of anything.

You don't need to change any settings because of Dessa. These habits come from the FTC's advice on AI voice scams, and they protect you and your family from voice scams of any kind:

  1. Don't trust the voice alone. If a call sounds like a family member in trouble, call that person back on a number you know is theirs.
  2. Check with someone else. If you can't reach them, contact another family member or one of their friends to confirm the story before you do anything.
  3. Watch how they want to be paid. A request to wire money, send cryptocurrency, or buy gift cards and read out the card numbers and PINs can be a sign of a scam. Scammers ask for these because the money is hard to get back.
  4. Report it. If you spot a scam call, report it to the FTC at ReportFraud.ftc.gov.

We'd add one more: agree on a private family code word that only you know, so anyone calling in an "emergency" can be asked for it. The next family dinner is a good time to pick one.

We'll update this if Block shares more about who can use Dessa, or when.

References