Most voice assistants work like a relay race. Your voice is turned into text, a language model writes a reply in text, and a speech engine reads that reply back to you. Block, the company behind Square and Cash App, thinks that relay throws away most of what makes a conversation human. On Friday its small AI research lab showed off Dessa, a voice model that skips text entirely.

The demos, shared by Block and its researchers on October 23, caught attention fast because of how human they sound, complete with laughs, pauses, sighs and the occasional awkward moment.

What Block announced

According to Block’s Dessa page, the model was trained from scratch on nothing but recordings of human conversation. There was no text and no pretrained language model in the mix. Block describes it as a true speech-to-speech model: audio goes in, audio comes out.

Dessa is also “full-duplex,” which means it can listen and talk at the same time, the way people do. Instead of waiting politely for you to finish a sentence, it can jump in, laugh mid-sentence or toss in quick acknowledgments like “yeah” and “got it.” The team says it learned when to do that simply by hearing how people really talk.

A few other details stand out:

  • Block says Dessa was trained on a single NVIDIA GB200 NVL72 rack, a modest footprint compared with the huge clusters used for frontier text models.
  • The company says it runs locally on a MacBook Pro with an M4 Max chip or better, rather than requiring a data center.
  • Block says inference runs faster than real time, so there is no long wait before it answers.
  • There is no public access yet. Block is offering early access to a small group of AI safety researchers.

TechCrunch reported that Dessa’s knowledge comes from the conversational data it was trained on, rather than from a separate language model bolted on for facts.

Why skipping text matters

Think about how much meaning lives in the way something is said. Sarcasm, a nervous laugh, a voice that cracks with sadness or a frustrated sigh can change a sentence completely. When speech is converted to text, most of that disappears.

Block argues this is more than an annoyance. Its announcement points out that today’s voice AI is nearly deaf to emotion, and gives the example of an AI that keeps chattering away while the person on the other end is clearly in distress. A model that can actually hear tone, the company says, is a step toward safer and more helpful assistants.

Many current voice tools add emotion with tags or instructions that tell the speech engine to sound happy or sad. Block says Dessa’s emotion is built in. TechCrunch noted it is not yet clear whether anything is steering that emotion behind the scenes.

Impressive, and a little uncanny

The sample clips include Dessa finishing a person’s sentence when they blank on a movie title, gasping at an anniversary surprise and playing a guessing game with a Block intern. In one demo it even chats with another company’s voice AI.

Not every moment is smooth. The Verge summed it up in its headline as an AI that talks to you like an awkward new friend. That mix of realism and oddness is exactly why Block is asking outside researchers to test it before any wider release.

The safety question

A voice that sounds this human raises obvious concerns. Realistic synthetic voices can be misused for scams, impersonation or manipulation, and a model that runs locally on a laptop is harder to monitor than a cloud service. TechCrunch suggested the safety-research program is likely aimed at exactly those situations where a voice agent could fool the person on the other end.

For now, Dessa is a research release, not a product you can download. That gives researchers time to probe the risks before the technology spreads.

What it could mean for everyday tech

Block sells tools to small businesses, and Square already uses voice AI to help with ordering. A capable voice model that runs on a device instead of in the cloud could eventually make voice ordering, customer support and personal assistants cheaper and more natural. It could also push rivals to rethink the text-in-the-middle approach most voice systems still use.

Whether Dessa itself becomes a product is an open question. Block has not announced pricing, a release date or plans for public access.

The bottom line

Dessa is an intriguing experiment: an AI that learned to hold a conversation by listening, much like a child does. It sounds strikingly human, runs on high-end consumer hardware and, for now, is only in the hands of safety researchers. If the approach holds up under scrutiny, the next generation of voice assistants may sound a lot less robotic, and that will make clear rules about disclosure and consent even more important.

Image: Music-technology-guitar-microphone-studio-amplifier-846852-pxhere by rafabendo via Wikimedia Commons, CC0 1.0