Dessa AI voice: Block unveils speech model trained only on conversation

Block, the payments and technology company led by Jack Dorsey, unveiled Dessa on Friday, 23 October 2026, an AI voice model that the company describes as trained entirely on recordings of human conversation, with no written text at all. Dessa is being released as a research preview: a small group of AI safety researchers has early access, and others can join a waiting list. The launch drew wide attention online over the weekend, with users praising how human it sounds and fraud watchers warning that such voices could make phone scams harder to spot.

Most voice assistants people use today work in steps. They first convert speech to text, then send that text to a language model, then turn the written reply back into synthetic speech. Each step adds delay and strips away much of what makes talk feel natural: tone, hesitation, laughter and the small sounds listeners make to show they are following, which linguists call backchannelling. A speech-to-speech model tries to skip the text stage entirely and work with sound from start to finish.

Full-duplex is the second idea behind Dessa. A walkie-talkie is half-duplex: only one side can talk at a time. A telephone call is full-duplex: both people can speak at once, interrupt and react mid-sentence. Block says Dessa behaves like the telephone, which is why early listeners described its demonstrations as unusually lifelike.

The training claim is the most unusual part. Large AI models are usually built on top of a "pretrained" base, a model that has already learned from huge amounts of text, and voice abilities are added later. Block says it did not do this and that Dessa learned only from recorded speech. If that holds up under outside testing, it would suggest a different route to conversational AI, one less dependent on written language. These are the company's claims; without published benchmarks or a model size, independent researchers cannot yet check them.

The local-running claim also matters. A model that runs on a high-end laptop does not need to send a user's voice to a remote server, which can help privacy and reduce delay.

Short-term: Only a small group of safety researchers can test Dessa for now, so ordinary users cannot try it. The immediate effect is on the debate: commentators have raised concerns that a voice this natural could be used to impersonate real people in phone fraud, and researchers will be looking at how Block limits misuse before any wider release.

Long-term: If Block's approach works as described, voice assistants could become far more conversational, which would affect customer-service call centres, accessibility tools for people who cannot easily type or read, language learning and in-car assistants. It would also add pressure on banks, telecom operators and regulators to strengthen caller verification, because "it sounded like a real person" would no longer be a useful test.

Who is affected: AI researchers and developers of voice products; businesses that run large call centres; people who rely on voice technology for accessibility; and phone users generally, including in India, where voice is a common way to use digital services. Block has not said which languages Dessa supports, an open question for a multilingual country.

| Feature | What Block says about Dessa |

| Input and output | Speech in, speech out, no text step |

| Conversation style | Full-duplex: listens and talks at once |

| Training data | Human conversation only, no text |

| Hardware | Runs on a MacBook Pro with M4 Max |

| Availability | Research preview for safety researchers, waiting list |

Questions and Answers