
Block unveils Dessa, an AI trained on human conversation
Block’s research lab has released a full‑duplex speech‑to‑speech model that learns entirely from raw audio, letting it converse in real time without converting to text.
24 Oct 2026 0 reads Compiled with AI
Block’s research lab announced the release of Dessa, a full‑duplex speech‑to‑speech AI that was trained from scratch using only human conversation. The launch came on October 23, 2026, and the lab published live demo recordings and opened a public waitlist. The announcement highlighted that no text or pretrained models were used and that the system can run locally on a MacBook Pro with an M4 Max or better chip.
Capabilities of Dessa
Unlike most voice assistants that convert speech to text and then generate a reply, Dessa receives audio and produces audio in real time. It listens and talks at the same time, using cues from the conversation to know when to back‑channel or to cut in. The model is designed to run faster than real time so that users do not experience noticeable delays. The lab also says it can run on consumer‑grade GPUs.

The demo clips show several conversational idiosyncrasies. In one recording, Dessa laughs mid‑sentence when something is funny, and in another it finishes a sentence when a speaker cannot recall a movie title. The model also reacts with emotion, gasping at a surprise and answering with what sounds like teary gratitude. In one recording it converses with GPT‑Live‑1, a third‑party AI that is not affiliated with Block. The recordings include subtitles that were added after the fact.
Block said the main motivations for developing Dessa were curiosity, better understanding between humans and AI, and safety. The company argued that voice AI that turns speech into text loses emotional nuance, which can be dangerous in situations such as a person in pain. To identify unforeseen risks, a small group of safety researchers was given early access, and a public waitlist is already open.
The AI research team behind Dessa is a small lab inside Block, founded by former members of the Toronto‑based startup Dessa. The team has built deep learning models since 2012 and generative speech systems for more than eight years. It reports directly to Jack Dorsey, Block’s co‑founder and chairman.