The uncanny valley just got a new resident, and it laughs at your jokes in real time. Block, the payments company behind Square and Cash App, has unveiled a voice model so expressive that the reactions split neatly into two camps: "wow" and "oh no."

Audio In, Audio Out, No Text in Between

Block's new model, Dessa, is a full-duplex speech model, meaning it can listen and talk at the same time, so it handles interruptions and laughter naturally. Block says it was trained entirely on human speech using a single Nvidia GPU rack, and that its knowledge comes from that conversational data rather than a separate language model, which speech systems usually rely on.

Demos show Dessa sounding tired after saying it had worked hard, sounding close to tears, and playing a guessing game with a staff member. It can also run locally on consumer hardware like a MacBook with an M4 Max chip or better. Block argues that converting speech to text loses information, calling that approach "ridiculous" and "potentially dangerous." Block's co-founder described talking to it as "like talking with a person." Competing tools such as ElevenLabs v4 and Fish Audio Drama 3 use voice tags to control emotion, while Block says Dessa's emotion is built into the model itself.

Great Bedside Manner, Terrible Scam Potential

A voice agent that sounds convincingly human, runs on a laptop, and needs no cloud is a massive engineering feat. It's also exactly what phone scammers dream about. Block hasn't released Dessa publicly and has asked outside safety researchers to test how a voice agent like this could deceive people.

The business angle is less dramatic but more interesting: Square already uses voice AI for ordering, and an on-device model could power more products at lower cost. Small businesses taking phone orders from an AI that sounds like a friendly regular? That's not sci-fi anymore. Running on-device also means lower latency and fewer recordings shipped off to someone else's servers, which is a genuine privacy plus. Keep in mind all of these claims are Block's own and haven't been independently verified.

Voice AI just crossed from "robotic but useful" to "wait, was that a person?" Your verification habits should cross with it. A family code word for "is this really you?" calls is starting to look less paranoid and more like basic hygiene.

Thinking about putting conversational AI in front of your customers, and want it to be helpful without being creepy? We help businesses plan AI integrations that earn trust.

Source: TechCrunch