On October 23, 2026, Jack Dorsey announced that a small research lab inside his fintech company Block had built something the voice AI industry has been circling for years: a conversation model named Dessa that learned everything it knows — language, facts, tone, laughter, sighs — from recorded human speech alone. No transcripts in the training pipeline. No pretrained language model underneath. No text anywhere. The team's own framing is unrestrained: "the most authentic full-duplex speech model ever built," with talking to it feeling "full of emotion, nuance, pauses, and a natural cadence." Strip away the marketing and what remains is still a genuine engineering provocation — because nearly every voice product you have used, from phone assistants to text-to-speech studios, runs on a text pipeline that throws away most of what your voice actually carries.
The official announcement post from Jack Dorsey's X account, October 23, 2026.
One disclaimer before anything else, because it matters more than the technology: Dessa is not available. There is no app, no API, no pricing page, no release date. Block has opened a waitlist and an early-access application aimed at a small group of AI safety researchers with institutional emails. Everything else on the page — and in this explainer — describes a research artifact being put under review, not a product you can wire into your stack this quarter. What follows is what it is, how it works, why the approach is different, and what it signals for people who make content for a living.
Who actually built this
The byline matters here. Dessa is not a startup chasing a launch cycle; it is the work of a deliberately small AI research lab inside Block, the parent company of Square, led by researcher Ragavan Thurairatnam, reporting directly to Dorsey himself. The name is a reunion: the lab's founders are former members of Dessa, the Toronto deep-learning startup Square acquired in 2020, and the team has been building deep learning models since 2012 and generative speech models for more than eight years. In other words, this is a veteran speech group with institutional patience and a CEO who personally funds contrarian bets — Block already runs voice AI in production for Square order-taking, and an on-device conversation model is an obvious long-term asset for a company whose business is small merchants and consumer payments.
The text pipeline problem, and why Dessa refuses it
To understand what makes Dessa different, look at the standard architecture first. Today's voice assistants and TTS systems are chains: automatic speech recognition converts your words into text, a language model thinks about the text, and a speech synthesis system converts the response back into audio. The chain works, scales, and ships — and it amputates the voice itself. Sarcasm, hesitation, a cracking note of sadness, the difference between a tired "fine" and a furious "fine" — none of that survives transcription into flat text. The most powerful AI systems in the world, as Block's page puts it bluntly, cannot hear how you feel.
The industry's workaround has been emotion tags — markup that instructs the synthesizer to sound excited or somber, the approach behind products like ElevenLabs' recent releases and comparable systems. It is a patch over the amputation. Dessa's bet is that if you train a model end-to-end on raw conversation, emotion stops being an annotation and becomes native behavior. The demonstrations on the official page are curated but revealing: the model laughs mid-sentence the moment something is funny, the way people do, instead of waiting its turn to perform a laugh; it gasps "oh my god" under its breath at a surprise and then answers with what sounds like teary gratitude; it plays a guessing game and wins; it backchannels — the little "yeah!", "got it!", "ooh" sounds that show a listener is tracking you; and when asked to speak up, it simply gets louder, unprompted by any tag. In one clip, it holds a live conversation with OpenAI's GPT-Live, both voices fully AI-generated, complete with a disclaimer that OpenAI is unaffiliated and has not endorsed Dessa.
The technical claims, as published
- Speech-only training. "No text, no pretrained models" — language was acquired from conversation recordings, and the team claims it learned from hundreds of times fewer words than text-based LLMs require.
- One-rack training. The entire model was trained on a single NVIDIA GB200 NVL72 — a fraction of the infrastructure behind frontier text models, which the team frames as proof the approach is tractable, not just elegant.
- Local inference. Dessa runs on a MacBook Pro with an M4 Max or better, faster than real time — no cloud round-trip, no awkward pauses in conversation.
- Full-duplex. It listens and speaks simultaneously, like a person, and knows when to cut in politely rather than waiting for silence.
The page also makes a research claim worth registering: that Dessa is the world's first AI with which you can hold a realistic live conversation that was trained only from human speech. Firsts like this are contestable, and other labs are racing down similar paths — treat "first" as the team's claim, not a settled fact. What is not contestable is that the artifact exists, runs on consumer hardware, and exhibits behaviors the field has been chasing.
Why this is hard: the case Block makes
The team's technical argument deserves more than a skim, because it explains why so few have attempted this. Speech never repeats: say the same sentence twice, as identically as you can, and the waveform data still differs — while typed text is byte-identical no matter who types it. A single spoken word consumes thousands of times more data than the same word in text. And the model must extract from that torrent not just vocabulary but everything that changes meaning: accents, pauses, breath, sarcasm, humor, intonation, timing, when to interrupt. Their worked example: "I didn't say she stole my CUDA code" means something different depending on which word you stress — seven readings, six of which plain text destroys. A baby learns to talk by listening alone; the team wanted to know if a model could too. Their answer: it can, "it's just very hard," and they claim to have found a way through — training the whole thing on one rack.
What it means for content creators — an honest read
Here is the professional translation, split into what changes today (nothing you can deploy) and what changes soon (a lot you should plan around).
- The credibility ceiling for synthetic voices is about to rise. The tell that betrays AI narration in most content — flattened affect, sentences pronounced rather than spoken — exists because of the text pipeline. If speech-native models prove out, the next generation of narration and dubbing tools will carry live emotional texture without manual tags, and the audible gap between a human narrator and a generated one will narrow to something only trained ears catch. If your brand's voice identity is being built today, assume the quality bar will jump a full generation within your planning horizon.
- Dialects stop being a data problem. Text models handle languages as written standards; your audience actually lives in spoken dialects that barely exist in writing. A model that learns from sound needs no dictionary of a dialect — just enough real recordings. Nothing in Block's materials promises multilingual support, and every published demo is English. But the engineering path to voices that handle regional Arabic, or any spoken-first language community, is now visibly open in a way it was not before.
- Voice agents get cheap enough to matter. A full conversation model that runs on a laptop changes the economics of voice-native products — customer-facing agents, interactive show notes, audience Q&A lines — from per-minute cloud billing to local compute. Block explicitly frames this as strategic for its own merchant products. For creators building audiences, voice agents are a coming distribution channel, and the cost curve just bent.
- Safety is the flip side of empathy. A voice that reads emotion is also a voice better equipped to influence it. The team acknowledges this directly — inviting outside safety researchers precisely because a model like this is new territory, and because an AI that cannot hear distress is its own hazard. Expect the safety discourse around synthetic voices to intensify before the creator tools even ship.
How it sits against tools you can actually use
| System | Input | Emotion | Public availability | Usable for production today? |
|---|---|---|---|---|
| Dessa (Block) | Raw speech only | Built into the model | No — safety-researcher early access | No |
| GPT-Live (OpenAI) | Internal text pipeline | Good, controllable | Yes, in ChatGPT | Conversations, not narration |
| ElevenLabs v4 and similar | Text with voice tags | Tag-driven, manual | Yes — apps and API | Standard for multilingual voiceover |
| Voxtral 3 (Mistral) | Transcription + TTS | Professional, even | Yes — commercial API | Ships Arabic dialects now |
For production work this quarter, the practical shelf remains the right-hand column: multilingual voiceover platforms, and API-grade speech systems such as Voxtral 3, which added Arabic dialect support at launch. We have also covered Google's Gemini TTS voice studio for Arabic narration workflows. Dessa belongs to a different shelf entirely: the one marked "next." Mixing them up is how roadmaps go wrong.
Limitations to keep in view
- No public access, no timeline. A waitlist and a researcher application form are the entire surface area. Any date you read elsewhere is speculation.
- English-only demos so far. No language roadmap has been announced. Treat any multilingual claim as inference, not fact.
- A conversation model is not a narration tool. Turning this architecture into TTS or dubbing products is a separate product-engineering effort that may never happen from Block itself — the company's stated interest is conversations, and its own Square merchant stack.
- Scale questions are open. One training rack is remarkable, but the model's knowledge comes from what it heard — the team itself frames how efficiently speech-only models learn facts and languages as an open research question.
- Demonstrations are demonstrations. Curated clips on a launch page are evidence of capability, not of reliability across accents, noise, adversarial users, or long sessions. That is precisely what the safety review is for.
Frequently asked questions
What is Dessa?
A full-duplex speech-to-speech conversation model built by an AI research lab inside Block, Jack Dorsey's company, announced October 23, 2026. It was trained exclusively on recorded human speech — no text, no pretrained models — and can run locally on a MacBook Pro M4 Max or better, faster than real time.
Can I try Dessa?
No. There is no public release. Block has opened a waitlist and an early-access application specifically for AI safety researchers using institutional emails. No consumer or API availability has been announced.
Does Dessa support Arabic or other languages?
Nothing official has been said about any language beyond the English-language demos. The speech-only training approach makes dialect learning theoretically more reachable, but that is analysis, not a company commitment.
How is this different from ElevenLabs or other voice tools?
Those products convert written text into speech and add emotion through manual tags. Dessa is a live conversation model with no text anywhere in its pipeline; emotion emerges from the model itself. One category produces narration you can ship today; the other is a research system you cannot yet touch.
Why should a content creator care about an unusable model?
Because it is a leading indicator. When speech-native training proves out, the narration, dubbing, and voice-agent tools you will buy next year inherit its properties: natural affect without tags, dialect handling from real recordings, and local-run economics. Teams building audio brands now should plan for a quality-bar reset, not a steady state.
The bottom line
Dessa is best understood as a direction, not a product: proof that a small team can train a full conversation model on speech alone, run it on a laptop, and make emotional intelligence native rather than tagged. For creators, the actionable reading is twofold. Short term, nothing changes — keep producing with the tools that ship. Medium term, the synthetic voice you deploy will likely descend from this architectural generation, and the gap between "sounds AI" and "sounds human" will keep collapsing. Build your audio workflows so the voice layer is swappable, because it is about to be swapped under you.
Until then, the durable half of any audio strategy is the script. ARWriter exists for exactly that half — long-form Arabic and English drafts written from your own sources, ready for whatever voice engine you hand them to. Good audio has always started as good text; that part is not changing.
Sources: Block's official Dessa page, the verified announcement post from Jack Dorsey's X account (October 23, 2026), and TechCrunch's coverage (October 24, 2026).
