By Hamza Faisal

Jack Dorsey's financial technology company Block has unveiled Dessa, an artificial intelligence voice model trained entirely on human conversation. No text, no pretrained models — just sounds, listened to until a machine learned how to hold a conversation.

Block announced Dessa on October 23, with Dorsey calling it "the most authentic full-duplex speech model ever built" in a post on X. "Talking with it feels like talking with a person," Block said, "full of emotion, nuance, pauses, and a natural cadence."

Trained on sounds, not sentences

Most voice assistants work in stages: they transcribe what you say into text, process that text, then turn an answer back into synthetic speech. Dessa skips the middle step entirely. It takes audio in and gives audio out, which Block says preserves the emotion, timing and tone that text transcription throws away.

"We trained Dessa from scratch on nothing but human conversation. Sounds! No text, no pretrained models. It learned to hold a real-time conversation simply by listening," Block said. The company added that large language models need an enormous amount of text, while Dessa learned language from hundreds of times fewer words.

Training on speech is harder than on text, Block said, because spoken words carry far more data and vary with every utterance. Even so, it trained Dessa on only a single NVIDIA GB200 NVL72 system, and the model runs locally on a MacBook Pro with an M4 Max chip or better, at faster than real time.

It talks back while you are still talking

Dessa is full-duplex, meaning it can listen and speak at the same time, much like a person in a real conversation. It can laugh mid-sentence, cut in, backchannel with a "yeah" or "got it," and change its volume on the fly. It recognises and expresses emotions including sarcasm, laughter, frustration and sadness, and decides for itself what to say, when to respond, and how to say it.

In Block's demonstrations, the model sounded tearfully joyful in one exchange and held a conversation with OpenAI's GPT-Live-1 in another. Demo coverage has called the results mostly natural, if occasionally awkward — a reminder that the public has seen company-chosen recordings, not an independent test.

Block frames that emotional hearing as a safety feature, not just polish. Most powerful AI models cannot hear how a user feels, the company argues, and a system blind to vocal distress could mishandle a moment when someone is in severe pain or trouble. Dessa, by contrast, is built to notice.

A research preview, not a product yet

For now, Dessa is a research preview behind a waitlist, not a product anyone can download. Early access is going to a small group of AI safety researchers, and Block has not published a parameter count, an API, a price, model weights or benchmark results. Project lead Ragavan Thurairatnam, whose small research lab inside Block built Dessa, has called it "the GPT-3 moment for voice."

That framing is deliberately ambitious: GPT-3 changed what people expected from text, and Block is betting a model that learns language the way children do — by listening long before reading — can do the same for voice. Whether Dessa holds up outside Block's own demos, or scales beyond conversation into real tasks, is still unproven.

Voice is becoming the interface companies want you to use, and the race is no longer just about understanding words. It is about sounding human — and Dessa is Block's claim that the way there is to stop teaching machines to read, and start letting them listen.