Block's small AI research lab — the company behind Square and Cash App, co-founded by Jack Dorsey — unveiled Dessa on October 23, 2026, a speech-to-speech conversation model in research preview. Its approach differs radically from today's assistants: instead of turning speech into text, reasoning over the text and then synthesizing a reply, Dessa takes audio in and returns audio, listening and speaking at the same time.
That mode is called full-duplex. In practice, according to Block, the model knows when to interrupt and when to simply signal that it is following the conversation, with short backchannels such as “yeah”, “got it” and “ooh”, and it hears — and produces — emotion, sarcasm, laughter, sighs and even a voice cracking. The company's argument is that voice carries information text discards, and that an AI unable to hear emotion might, for example, fail to notice that someone is in pain.
The most striking point is the training. Block says it trained Dessa from scratch using only human conversation, with no text and no pretrained models — like a child who learns to talk by listening, without ever having seen a written word. The company also says the model learned from “hundreds of times fewer words” than large language models and that training used a single NVIDIA GB200 NVL72 system. These are statements from Block, not yet independently audited: the company has published no parameter count, dataset size, latency figures or formal benchmarks.
Also according to Block, inference runs faster than real time and locally on a MacBook Pro with an M4 Max chip or better. The lab is small, was founded by former members of the Toronto startup Dessa, is led by Ragavan Thurairatnam and reports directly to Jack Dorsey. Thurairatnam described the launch as a “GPT-3 moment for speech” — an opinion, not a measured result.
Instead of benchmarks, Block released recordings it describes as live and unsteered, with subtitles added afterwards. In them, Dessa laughs mid-sentence, raises its volume when asked, finishes the sentence of someone who forgot a movie title, and reacts with surprise to an announcement. One demo has it talking to GPT-Live-1, a third-party model that, according to Block itself, is not affiliated with the company and has not reviewed or endorsed Dessa.
Access, for now, is restricted. A small group of AI safety researchers will get early access, signing up with an institutional email, precisely to map risks before a wider opening — and Block says it wants “outside eyes” on new territory. For the general public, a waitlist is open, with no launch date, pricing or open weights announced. The company itself hedges its main claim: it says that, “as far as it knows”, Dessa is the first AI you can have a realistic live conversation with that was trained only on human speech.
Image credit: illustration from AI Chat Daily's coverage of Dessa.
Sources
- Block — página oficial da Dessa
- AI Chat Daily — Block unveils Dessa, a full-duplex voice model trained only on human speech
- RuntimeWire — Block unveils Dessa, a speech-only AI model built for real-time laptop use
- Julian Goldie — Dessa AI: Block's Speech-Only Voice Model Explained
Transparency: this article was produced with AI assistance, based on cross-checking public posts on X, as reported by news outlets, and the internet sources listed below, including Block's official page. Claims about training and performance are company statements, not yet independently verified. Check the original sources before making decisions.
Muse
