Dessa AI Voice Model: Block’s Speech AI That Hears Emotion

The Dessa AI voice model is Block’s first big step into speech AI. Jack Dorsey’s company unveiled it on October 23, 2026, as a research preview. Block says the model learned to talk only by listening to people, with no text at all. You cannot use it yet, though.

Dessa comes from a small AI research lab inside Block, the company behind Square and Cash App. Block calls it a full-duplex speech-to-speech model, trained from scratch on human conversation.

The company went big on the claim. It calls Dessa “the most authentic full-duplex speech model ever built,” as reported by TechCrunch. However, Block has not published independent benchmarks to back that up.

Instead, the launch rests on recorded demos. In them, Dessa laughs, sighs, pauses and even sounds tired, like a person at the end of a long day.

Dorsey framed the project in one line on X. “We believe voice is the most natural way to interact with intelligence,” he wrote.

Block’s own pitch is blunter. “Interacting with current voice AI can just plain suck,” the company wrote on its Dessa page. In its view, today’s strongest AI systems are nearly deaf to the emotion in our voices.

There is a business angle too. Square already uses voice AI for ordering, and TechCrunch notes that a model running on the device could make such products cheaper. For now, though, Block lists curiosity, better human-AI understanding and safety as its goals.

Most voice assistants work in turns. First they turn your speech into text, then a language model writes a reply, and finally a voice reads it out. As a result, tone, sarcasm and hesitation get lost on the way.

Dessa skips the text step. It hears audio and answers in audio, which makes it speech to speech AI in the plain sense of the term.

It is also full duplex. That means it listens while it talks, so it can be interrupted and can interject. In addition, it drops in small cues such as “yeah!” or “got it!” while you speak, the way people do on a phone call.

Block shared more about how it built Dessa than about its plans for it. Here are the confirmed details from Block and the launch coverage.

The hardware line stands out. A model that trains on one rack and runs on a laptop is small by today’s standards. Even so, that efficiency is Block’s own claim until outside researchers test it.

Dessa is not the first voice AI with feelings. OpenAI released GPT-Live, a full-duplex model, earlier in 2026, and Block’s demos include a recorded talk between Dessa and GPT-Live-1.

The difference is in how emotion gets there. TechCrunch points out that tools like ElevenLabs v4 and Fish Audio Drama 3 use voice tags, small written cues that tell the voice to laugh or whisper. Block says Dessa’s emotion is built into the model instead.

That matters if you make voiceovers today. Our ElevenLabs review covers how those tags and voices work in practice, and you can actually sign up for it.

The name is not new. Dessa was a Toronto machine learning startup, founded in 2016 as DeepLearni.ng. Square bought it on February 7, 2020, according to BetaKit, and kept the Dessa brand in Toronto.

Ragavan Thurairatnam was Dessa’s head of machine learning back then. Today he leads the lab, which reports to Dorsey and lists six researchers and three interns.

According to Block, the team has built deep learning models since 2012 and speech models for more than eight years. Meanwhile, Thurairatnam called the launch “a GPT3 moment for speech” in a post on X.

Not really. Block opened a waitlist, but it asks for an institutional email from a university, company or lab. Early access goes to a small group of AI safety researchers first.

There is also no API and no price. People searching for Dessa on Hugging Face or for a Dessa TTS download will come up empty for now, since Block has not released the model files.

Block says the safety work is about deception. It wants researchers to probe cases where a voice agent could fool people. Meanwhile, the demos already show limits: Gizmodo noted that Dessa kept guessing in a game it had already won.

Dessa is still behind a waitlist, but other natural-sounding voices are open today. Compare speed and pricing in our Cartesia AI review, a fast voice API built for real-time agents.

Sources

Block, Dessa, TechCrunch (October 24, 2026), Gizmodo (October 24, 2026), Let’s Data Science (October 23, 2026), BetaKit (February 7, 2020). *Photos: Jack Dorsey by Steve Jurvetson, CC BY 2.0, resized and placed on a blurred background; Microphone in SPNN Podcast studio by Myotus, CC BY 4.0, cropped and resized.*