Situation report: Voice AI has a hearing problem, and Jack Dorsey’s Block says it has a fix.
Block, the company behind Square and Cash App, has built Dessa, an AI model that learns to talk straight from audio and never reads text along the way, according to The Verge AI. The Verge’s headline describes the early result well: it talks to you “like an awkward new friend.” It’s not polished yet. Still, the idea behind it is a big one.
🎯 What Block Launched
Block’s starting point is simple. In its own words: “Babies learn to talk without ever seeing a word written down. We wanted to see if a model could too.” Its answer: “It can. It’s just very hard.”
What Dessa is, based on what’s been reported so far:
- Speech-native model. It learns speech directly. It doesn’t learn written language first and add a voice later.
- No text bottleneck. It skips the step where your speech gets turned into a transcript before the AI responds.
- Built for human timing. Block says today’s voice AI “can’t keep up with human timing.” Dessa is meant to handle the pauses, overlaps and rhythm of real conversation.
- Hears emotion. The main goal is picking up how you say something, not just the words you use.
🔍 How It Differs From Today’s Voice AI
Most voice assistants still run on a relay. Your speech becomes text. A language model reads that text. Then a separate engine turns the reply back into audio. It works, but you lose a lot along the way.
Tone, hesitation, a shaky voice, sarcasm: none of that survives being turned into a transcript. Block says current systems “flatten what we say into text.” The industry has been moving toward speech-to-speech models for a while, but most still lean on text somewhere in the process. What stands out about Dessa is that Block went all in on audio from the very first step of training.
⚠️ The Safety Angle
Block’s argument isn’t mainly about convenience. It’s about safety.
“Speech is the most natural interface to any intelligence, super or otherwise, and today’s voice AI wastes it,” the company says. As AI gets more capable and more independent, Block warns it is “still nearly deaf to the emotion in our voices.”
Then comes the example that hits hardest:
“Imagine being in serious pain while a voice AI has no idea and just keeps going. That’s the safety problem. Dessa is a step toward fixing it.”
This matters because voice agents are moving into customer service, healthcare intake and personal assistants. An AI that can’t tell a calm caller from a panicking one is a real risk in those jobs.
🛠️ Where This Could Land
The report doesn’t list specific products. Based on Block’s framing, though, the obvious places to look are:
- Customer support. Spotting frustration before a call goes badly. Block runs huge payment and support operations of its own.
- Health and wellness. Catching distress that a transcript would miss.
- Accessibility. More natural conversation for people who depend on voice.
- AI companions and assistants. Talking with AI that feels less robotic and more like an actual exchange.
🚧 Caveats
- It’s awkward. The “awkward new friend” line is honest. Dessa’s conversation skills are still rough.
- “Very hard” is Block’s own phrase. Learning speech without text is a tough research problem, and Block admits that.
- No availability details yet. The source doesn’t say anything about pricing, API access or a public release date. Treat Dessa as a research milestone for now, not a product you can use.
- Unusual source. Block is a payments company, not a frontier AI lab. That makes this interesting, but it also means it has to prove itself against the big labs’ voice models.
📡 Outlook
Block is betting that the next big step in voice AI is a model that actually hears you, not one that’s simply smarter. If Dessa’s approach scales, the text-in-the-middle setup most voice assistants use today could start to look outdated. Watch for benchmarks, a developer release, and whether the larger labs answer with their own speech-native models.
You can find the full story at The Verge AI.