Dessa: Block's speech model trained only on listening

Block, the payments company led by Jack Dorsey, announced Dessa, a full-duplex speech-to-speech model it calls "the most authentic full-duplex speech model ever built," trained only on human speech with no text pretraining. Block claims Dessa handles simultaneous listening and talking, including interruptions, and is available as an app with a waitlist for general access, but has published no weights, model card, parameter count or license. Block's announcement, dated October 23, 2026, positions Dessa for its Square seller and Cash App phone-support businesses, while the open-weights version of the capability remains unshipped.

Jack Dorsey’s Block just crossed from payments into the conversation business. The announcement: Dessa , “the most authentic full-duplex speech model ever built,” and the demo line that matters: everything it knows came from human speech. No text pretraining story, no transcribe-then-reason pipeline - audio in, audio out, one model, and it listens while it talks, the way people do. What Dessa is TLDR: Dessa is a full-duplex speech-to-speech model from Block that talks like a person: emotion, nuance, pauses, natural cadence. It enables realistic live conversation with a machine - interruptions and all. The difference between Dessa and a regular voice AI is the pipeline: current voice assistants transcribe your voice into text throwing away how you said it , reason in text, then speak text back. Dessa skips all three hops - it heard you, and it answers with tone intact. The page’s own explanation of why the category has sucked: “Interacting with current voice AI can just plain suck. It turns your voice into text and throws away an incredible amount of information text alone can’t carry. The most powerful AIs in the world can’t hear how you feel. We think that’s ridiculous, leads to a worse experience, and is potentially dangerous.” The claims, sorted Real and checkable: full-duplex simultaneous listen + talk - barge-in works, the thing most voice stacks fake with turn-taking heuristics , speech-only training claimed a first for a production-quality conversationalist - “as far as we know” is doing heavy lifting in their sentence, and they know it , app availability Download today + waitlist for general access . Not yet real: weights none published - no model card, no parameter count, no license , the local angle this is an app-preview, not a research release , and the “most authentic” superlative marketing until someone measures it - though the demo transcripts, with mid-sentence interruptions and emotional reads, are the strongest signal so far . Why this category keeps failing - and why this try is different Every voice assistant you have used is three models stapled together: speech-to-text, an LLM, text-to-speech. Each hop loses information sarcasm dies in transcription; pauses die in synthesis and each adds latency the robocall cadence . The full-duplex approach - one model, simultaneous, speech-native - is where the field has been heading since the duplex research line Moshi’s early full-duplex work, then the speech-native experiments . Dessa is the first CONSUMER-positioned release from a major fintech company betting that approach is ready. The fit for Block: their businesses Square sellers, Cash App run on phone support and point-of-sale conversations. A voice model that hears frustration instead of transcribing “I guess that’s fine then” is worth real money to them. This is a company-grade product motivation, not a research demo. For the site’s category Dessa is a closed-preview speech model - tracked, not a modeldex row no weights, no API card . The speech-models ladder on the site: Voxtral 3 still waits for its weights, Gradium’s Phonon is microcontroller-tier, and Dessa stakes out the top of the interaction-quality axis: not smallest, not openest - most human-feeling. The open-weights version of this capability remains THE open problem; nobody has shipped it yet. Note: if the waitlist opens or weights surface, the page gets re-verified and this piece gets the update. Until then: the most interesting conversation model is the one you cannot run locally. Sources: Block’s announcement block.xyz/dessa , the x.com/jack post Oct 23, 2026 . Related on this site: speech-model rows live in the modeldex; the engine catalog covers the local runtimes this category eventually lands on.