Block presents Dessa, a full-duplex speech model trained for conversation – Unite.AI

Block’s in-house AI research lab announced on October 23, 2026, Dessa, a full-duplex speech synthesis model that the company says was trained from the ground up on nothing but human conversations, releasing live demo recordings, a public waitlist, and early access to a small group of AI security researchers.

The ad paired Block’s Dessa page with a post on X by Ragavan Thurairatnam, the lab’s head. Thurairatnam described Dessa as the first AI model capable of holding a realistic conversation after training entirely based on speech, said it can run on consumer GPUs, and called the release “a GPT3 moment for speech.” On his page, Block writes that, as far as the company knows, Dessa is the world’s first AI capable of holding a realistic live conversation after being trained only by human speech.

Block describes Dessa as a true model of speech synthesis: audio in, audio out. It’s full-duplex, so it listens and speaks at the same time, and the company says it learned from a real conversation when to intervene and when to backchannel to show it’s following. Block says the model natively feels and speaks emotions, sarcasm, laughter, a frustrated sigh and a voice cracking with sadness; that inference is faster than real time, so answers come without awkward pauses; and that Dessa runs locally on a MacBook Pro M4 Max or better. The company calls it the most authentic full-duplex voice model ever built and says talking to it is like talking to a person.

A training approach based exclusively on speech

According to the page, Dessa was trained from scratch on sounds alone, without text or pre-trained models, and learned to converse in real time solely by listening. Block argues that the difficulty of this task is easy to underestimate because people learn to speak so easily. Speech data is never repeated: even saying the same sentence twice in exactly the same way produces different data, while the typed text is identical regardless of who is typing it. In terms of data, the page states, a spoken word takes up thousands of times more bytes than its written form, and the position of the accent changes meaning; an example sentence on the page takes on a different meaning each time a different word is underlined.

Block says that Dessa learned the language from hundreds of times fewer words than large language models, and that the team figured out how to train it on a single NVIDIA GB200 NVL72 system. The page lists the open research questions created by the approach: how efficiently such models learn language and facts over text, how efficiently they can learn new languages, and what kind of in-context learning they can perform.

Recorded demonstrations

It blocks posted clips that it describes as live recordings of people talking to Dessa. The company says that no one guides the model during these conversations and that Dessa herself chooses what to say, when to say it and how to say it, including her own voice and personality; subtitles were added after the fact.

Moments described include Dessa laughing in the middle of a sentence when something is funny, rather than after the speaker has finished; raise your voice when asked to speak; and finishing a person’s sentence when they leave a movie title blank. In another clip, Dessa gasps softly at a surprise anniversary announcement before responding with what the page describes as tearful gratitude, and a longer recording of life advice is described as genuinely empathetic. A recording shows Dessa conversing with GPT-Live-1, a third-party AI model; Block notes that both voices in that recording are AI-generated, and that the developer of GPT-Live-1 is not affiliated with Block and has not reviewed or endorsed Dessa.

Security, access and laboratory

The team gives three reasons for building Dessa: curiosity, better understanding between humans and AI, and security. The page notes that children learn to speak without ever seeing a written word, and says the team wanted to see if a model could do the same. As for safety, Block argues that the AI ​​continues to gain power and independence while remaining incapable of registering emotion in the speaker’s voice, giving the example of a person in severe pain while the vocal AI has no idea and continues to move forward. The company describes Dessa as a step towards solving this problem and accuses current voice AI of reducing speech to text and losing information carried by sound alone.

Block is offering a small group of AI security researchers early access to spot risks the team didn’t anticipate, and its application form asks for a name and institutional email from a university, company, or lab, with an optional link to Google Scholar or the blog. A public waitlist is open, more registrations will be shared on Thurairatnam’s X account, and the page says launch links are coming soon.

Dessa comes from what Block describes as a very small AI research lab within the company, founded by former members of Dessa, the Toronto deep learning startup. The team says it has been building deep learning models since 2012 and generative speech models for more than eight years, and reports directly to Jack Dorsey, Block’s co-founder, president and chief operating officer. Thurairatnam’s biography X states that he previously founded and sold Dessa and lectured at the University of Toronto. Alongside him, the page lists researchers Louay Hazami, Hashiam Kadhim, Qi (Alex) Liu, Rayhane Mama, Aydin Polat and Marc Tyndel; interns Cara Ji, Jibran Iqbal Shah and Dvir Zagury-Grynbaum; and former interns Aryan Sharma and Olivia Yan.