Block, the company behind Square and Cash App, has published recordings of Dessa, a voice AI that it says learned to talk from human conversation alone: no text, and no existing language model underneath. It is not a product. There is a waitlist, and AI safety researchers get early access first.
A transcript keeps the words and drops the person
Most phone bots are built as a chain. One piece turns your voice into text, a second reads the text and writes a reply, and a third reads that reply aloud. By the time the second piece starts work, the sigh, the pause and the flat voice are gone. "That's fine" looks the same on the page whether it was said warmly or through clenched teeth.
Newer models such as OpenAI's GPT-Live-1 listen to the audio directly, but they hand the real thinking to a text model, so what they know comes from reading. Dessa's knowledge came from listening only. That means tone is part of how it understands words from the start, not something added afterwards.
Call-centre software can already score a caller's voice for stress and warn a supervisor, but it works as an alarm beside the call. A model that hears the tone is also the one answering, so it can change what it does while the caller is still talking.
Why sound is hard to learn from, and why this one was cheap
Sound is much heavier material than text. A spoken word takes thousands of times more data than the same word typed, and nobody says a sentence the same way twice. A model also has to learn timing: when to cut in, and when to say "yeah" to show it is listening. Block says Dessa learned language from hundreds of times fewer words than text models read. It trained Dessa on a single rack of 72 NVIDIA chips, a small fraction of what the biggest text models use.
If that holds up outside Block, the listening-and-speaking layer of voice AI stops being something only a few labs can build. Block describes Dessa as small enough to run on a high-end MacBook Pro. For a clinic, a model like that would mean calls handled without the audio leaving the building.
The thinking still happens elsewhere. Dessa chats, laughs and helps a person remember a movie title, but nothing Block has shown books an appointment or looks up an account. The product people will actually use pairs a voice like this with a text model that does the lookups, the way ChatGPT Voice already reaches email, calendar and Slack.
What a clinic's phone line looks like with an ear
Picture a dental clinic that gets 50 calls a day. Say 40 are routine: a booking, a change of time, a question about opening hours. At three minutes each, that is two hours of receptionist time a day. The other ten are people upset about a bill or a cancelled slot, and those are the calls that need a person.
A bot that cannot hear tone treats all 50 the same, or sends callers to a human only after they say the magic words. A bot that hears the strain in a voice can finish the 40 routine calls and pass the upset ones to the receptionist with a note, before the caller has to ask. The staff lose the easy calls and keep the hard ones, which is where a person adds the most.
A human-sounding voice no longer proves anyone is there
The same skills make a fake caller convincing. A voice that laughs mid-sentence, sounds tired and waits for the right moment to speak is what a real person sounds like on the phone. ElevenLabs can already clone a voice from 10 seconds of audio, and a live model that handles timing and feeling removes the stiff pause that used to give bots away.
So any request for money or a password over the phone needs a check that does not depend on how the voice sounds. Hang up and call back on a number you already have, and agree a word with your family or your staff that a stranger could not guess.