Block Unveils Dessa, a Voice Model That Learned to Talk from Speech Alone

What happened

The launch includes recorded live demos of people talking with Dessa: it laughs mid-sentence when something is funny, gets louder when asked to speak up, and finishes a person's sentence when they blank on a movie title. Another clip shows Dessa talking with GPT-Live-1, a third-party AI model; Block notes that both voices are AI-generated and that GPT-Live-1's developer is not affiliated with Block.

The training approach is the core of the announcement: Block writes on the launch page that Dessa was trained from scratch on human conversation audio alone, with no text data or pretrained models, and that as far as the company knows, it is the world's first AI able to hold realistic live conversations after learning only by listening to human speech. The page adds the human parallel the team started from: babies learn to talk without ever seeing a written word, and the team wanted to see whether a model could do the same.

Project lead Ragavan Thurairatnam announced the release on X, calling it a "GPT-3 moment for speech"; Jack Dorsey introduced the product the same day. Block opened a public waitlist and granted early access to a small group of AI safety researchers, asking them to find risks the team has not anticipated.

Key facts

Full-duplex: Dessa listens while it talks, and the company says it learned from real conversation when to cut in and when to backchannel to show it is following along; it natively hears emotion, sarcasm, laughter, a frustrated sigh and a cracking voice; inference runs faster than real time, so replies come without an awkward pause.

Local execution: it runs locally on a MacBook Pro M4 Max or better, with no cloud inference needed. Block says training used a single NVIDIA GB200 NVL72 system, and that the model learned language from hundreds of times fewer words than large language models (both company-reported).

Open questions: the launch page lists the research questions this approach raises — how efficiently such models learn language and facts compared with text models, how efficiently they learn new languages, and what kinds of in-context learning they can do. Block also concedes the evidence so far is recorded demos with no public verifiable benchmarks; real-world reliability and safety await the early-access safety review.

Background

Dessa comes from a very small AI research lab inside Block, founded by former members of Dessa, the Toronto deep-learning startup — Ragavan Thurairatnam founded and sold that company and has lectured at the University of Toronto; the team says it has built deep learning models since 2012 and generative speech models for more than eight years, and it reports directly to Block cofounder Jack Dorsey.

The company gives three reasons for building it: curiosity, better understanding between humans and AI, and safety — AI keeps gaining capability and independence while remaining unable to register the emotion in a speaker's voice, such as a person in serious pain while a voice AI has no idea and keeps going; the company blames current voice AI's habit of reducing speech to text first, losing the information only sound carries.