Block Dessa AI Voice Model Explained: Jack Dorsey's Full-Duplex Speech AI Trained Only on Sound (USA)
Image: Block — Dessa (official brand visual)
📌 KEY HIGHLIGHTS
- Block’s in-house AI lab unveiled Dessa on 23 October 2026 — announced by Jack Dorsey on X and on the official page block.xyz/dessa.
- Dessa is full-duplex speech-to-speech: audio in, audio out, listening and speaking at the same time, with no transcription step in between.
- It was trained from scratch on nothing but human conversation — no text, no pretrained models — learning language “hundreds of times fewer words” than large language models see, per Block.
- Training ran on a single NVIDIA GB200 NVL72, and inference runs locally on a MacBook Pro M4 Max or better, faster than real time, per the company.
- It is a research preview behind a public waitlist; early access goes first to a small group of AI safety researchers. No parameter count, API, price or weights announced.
What happened?
On 23 October 2026, Block — Jack Dorsey’s fintech company behind Square and Cash App — pulled the curtain on Dessa, a new AI voice model from a very small AI research lab inside the company. Dorsey announced it on X with a short post: “Voice is the most natural way to interact with intelligence and Dessa is the most authentic full-duplex speech model,” he wrote, calling the model compact enough to run on a MacBook.
The lab itself is worth knowing about: Block says it was founded by former members of Dessa, a Toronto deep-learning startup — the model inherits the startup’s name. The team has been building deep-learning models since 2012 and generative speech models for over eight years, and it reports directly to Dorsey. The lab’s lead, Ragavan Thurairatnam, called the release “a GPT-3 moment for voice” (reported by PANews), and the official announcement page lists the full research team — six researchers, three interns and two former interns, in alphabetical order.
The announcement landed with live demo recordings on block.xyz/dessa and immediately drew coverage on 23–24 October from the Indian Express, The Tribune, Asianet, Unite.AI and others. The one-line pitch: a model that learned to talk the way a baby does — by listening, without ever reading a word.
How the Block Dessa AI voice model works
Almost every voice assistant you have used works in three stages: your speech is transcribed to text, a language model writes a text answer, and a text-to-speech voice reads it back. Dessa collapses all three stages into one model: sound goes in, sound comes out, and there is no text anywhere in the middle.
That design choice is the whole point. Text strips away tone, pacing, hesitation, laughter and timing — the information that tells you how someone means what they say. Block argues that today’s voice AI is stilted and slow precisely because it flattens conversation into text, and that “the most powerful AIs in the world can’t hear how you feel” — which the company calls not just a worse experience but potentially dangerous (imagine a voice assistant that cannot hear that you are in serious pain and just keeps going).
Training a model on raw speech is, by Block’s own telling, brutally hard. A single spoken word takes thousands of times more bytes than the same word in text; the same sentence never comes out acoustically identical twice; and accents, pauses, breathing, sarcasm, humour and inflection all change meaning. Block claims it solved this and trained Dessa on only one NVIDIA GB200 NVL72 — far less compute than frontier text-model runs.
The result, according to the demos: Dessa laughs mid-sentence when something is funny (other models wait for you to finish, then laugh on cue), finishes your sentence when you blank on a movie title, speaks louder when asked, gasps “oh my god” under its breath and answers with what Block describes as teary gratitude, and even holds a conversation with a third-party voice model (GPT-Live-1, which Block notes is unaffiliated). Crucially, nobody steers it in the recordings — Block says Dessa picks what to say, when to say it and how, down to its own voice and personality.
Image: Wikimedia Commons / Dejan Krsmanovic, CC BY 2.0 — Illustrative image: a studio condenser microphone close-up (Dessa is audio-only software; Block has released no hardware or interface photos).
Why it matters for USA readers
Dessa is a US-built frontier AI bet — and it arrives at a moment when the voice-AI race is heating up. The same week, Mistral launched Voxtral 3, and OpenAI has been pushing its voice features hard. What makes Block’s approach different is architectural: if a model can genuinely learn language from sound alone, it challenges the near-universal assumption that massive text corpora are a necessary scaffold for language competence. For American developers and researchers, that opens a new research direction — and for everyone else, it promises voice assistants that actually hear you.
There is a practical upside too. Because Dessa runs locally on a MacBook Pro M4 Max, it points toward voice AI that does not send your conversations to a server — better privacy, lower latency and no per-minute API bill. For Indian readers, the longer-term prize is multilingual voice AI: accents, code-switching and the natural rhythm of Indian-English conversation are exactly the things text pipelines handle worst.
Should you care? (Verdict)
Dessa vs typical voice assistants
| Feature | Dessa (Block) | Typical voice assistant |
|---|---|---|
| How it works | One speech-to-speech model: audio in, audio out | Speech → text → model → text → speech |
| Listening while talking | Yes — full-duplex, you can interrupt | Usually takes turns; waits for you to finish |
| Trained on | Human conversation audio only; no text | Mostly text, plus a separate voice layer |
| Pauses, laughter, emotion | Learned natively from real speech | Often lost or added artificially |
| Training compute | One NVIDIA GB200 NVL72 (per Block) | Large multi-datacenter GPU clusters |
| Runs on | MacBook Pro M4 Max or better, locally (per Block) | Usually in the cloud |
| Available now | Waitlist only; safety researchers first | Widely available |
| Knowledge & tools | Unproven; facts come “from raw sound” | Vast text knowledge, tool use |
FAQs
What is Block’s Dessa AI voice model?
Dessa is a full-duplex speech-to-speech AI model unveiled by Jack Dorsey’s Block on 23 October 2026. It was trained from scratch on human conversation audio alone — no text, no pretrained models — and can listen and speak at the same time, natively hearing and expressing emotion, sarcasm and laughter. It is currently a research preview behind a public waitlist.
How does Dessa work without any text training?
Dessa skips the transcription step entirely: raw audio goes into a single model and raw audio comes out. Because it never converts speech to text, it keeps the paralinguistic information — tone, pacing, hesitation, laughter — that text pipelines discard. Block says it learned language from “hundreds of times fewer words” than text models, the way a baby learns to talk by listening.
When can I try Dessa? How does the waitlist work?
The public waitlist is on Block’s official Dessa page (block.xyz/dessa). Block says early access is going first to a small group of AI safety researchers “to find the risks we haven’t anticipated” — the company calls a model trained only on sound “new territory.” No public release date, pricing, API or open weights have been announced.
What are Dessa’s limits right now?
So far the only evidence of Dessa’s abilities is Block’s own demo recordings; there is no published evaluation, no independent testing, no parameter count and no benchmark scores. Independent coverage has noted the demos show limitations — misidentifying a speaker, unexpectedly changing topics — and text-trained models still know far more facts and can use tools, which Dessa has not demonstrated.
Takeaway: Dessa is Block’s boldest AI claim yet — a voice model that learned language by listening alone. The demos are impressive, the science is genuinely interesting, but every impressive claim still rests on company-selected recordings. Until independent researchers get their hands on it, treat Dessa as a fascinating preview, not a product.
Follow USA Tech on TechByCountry for tomorrow’s 6 PM update.
📸 Get tech in the last 24 hours on Instagram — follow @_techinlast24hrs for daily posts (9 AM–9 PM) and reels at 12:30 PM & 8:30 PM.
Sources
- Block — “Dessa - the Conversation Model trained only on speech” (official announcement page, 23 Oct 2026; also the source of the featured brand visual)
- Jack Dorsey (@jack) on X — Dessa announcement post, 23 Oct 2026 (quoted in the Indian Express)
- Indian Express — “Jack Dorsey’s Block unveils new voice AI model that talks like a real human” (24 Oct 2026)
- The Tribune (ANI) — “Jack Dorsey’s Block launches Dessa; AI voice model trained entirely on ‘human conversation’” (24 Oct 2026)
- Asianet Newsable (ANI) — “Block Unveils Dessa: An AI Voice Model Trained Only on Human Speech” (24 Oct 2026)
- Unite.AI — “Block Unveils Dessa, a Full-Duplex Speech Model Trained on Conversation” (23 Oct 2026)
- Crypto Briefing — “Block unveils Dessa, an AI voice model trained entirely on human conversations” (23 Oct 2026; demo limitations)
- Related reading on TechByCountry: Kimi K3.1 Pro, the open-weight model sitting near the frontier, the MiniMax M3.1 Flash open-weights release, and Anthropic’s Claude Duet proactive agent
- Image pages: Block’s official Dessa page (featured brand visual); studio condenser microphone close-up, Wikimedia Commons / Dejan Krsmanovic, CC BY 2.0
.jpg){kind=link}
Comments
Post a Comment