What if a voice assistant never read a single word? Block Dessa, unveiled on October 23, 2026, by Jack Dorsey’s company Block, is a full-duplex speech model that Block says learned to hold a live conversation from recorded human conversation alone.
No text. No pretrained model underneath. That is the headline claim — and it matters for voice builders, even though you cannot use Dessa today.
What Block Actually Announced
Block’s small research lab published a project page, not a research paper. It shared recorded conversations, not benchmark tables.
The lab says each clip is a live, unsteered recording, and Dessa picks what to say, when, and how. Subtitles were added afterward.
Dessa at a Glance — From Block’s Official Page
| Maker | Block’s small AI research lab |
| Announced | October 23, 2026 |
| Type | Full-duplex, speech-to-speech model |
| Training data | Conversation audio only, no text (per Block) |
| Training hardware | One NVIDIA GB200 NVL72 (per Block) |
| Runs on | MacBook Pro M4 Max or better (per Block) |
| Inference | Faster than real time (per Block) |
| Public access | None announced — safety researchers only |
| Benchmarks | None published |
Source: block.xyz/dessa. All claims are Block’s.
The Evidence: What Block Did Publish
No parameter count, no latency figure, no benchmark score — but striking claims, quoted exactly:
- “Hundreds of times fewer words” — Dessa learned language from hundreds of times fewer words than LLMs need, per Block.
- “One NVIDIA GB200 NVL72” — Block says that single system trained the model.
- “Over eight years” — the lab says it has built generative speech models for over eight years, and deep learning models since 2012.
- “Faster than real time” — Block says inference runs faster than real time, so there is no awkward wait.
How Dessa Differs From Today’s Voice AI
Most voice assistants work in turns: you speak, it transcribes to text, a model replies in text, and a voice reads it aloud. Block’s argument: that pipeline throws away timing, tone, laughter, and emotion.
Full-duplex means the model listens and talks at the same time, like a phone call — not a walkie-talkie where each side waits its turn. Block says Dessa knows when to cut in and backchannel (“yeah!”, “got it!”, “ooh”), and natively hears and speaks emotion, sarcasm, laughter, sighs, and a cracking voice.
Today’s Voice Pipeline vs. Dessa (Per Block)
Today’s voice AI
1. You speak
2. Speech becomes text
3. A text model writes a reply
4. A voice reads it aloud
Timing and emotion lost at step 2.
Dessa (Block’s description)
1. You speak — audio goes straight in
2. Dessa listens while it talks
3. Audio answer, with laughter, backchannels
Nothing flattened into text.
Per Block.
How It Works, Step by Step
- You speak. Audio goes straight in. No transcription, no text model.
- Dessa listens while it talks. It can be interrupted, can interrupt, and backchannels while you speak.
- It replies in audio. Dessa picks its words, timing, voice, and personality, per Block.
- It runs locally. Per Block: MacBook Pro M4 Max or better, faster than real time.
Why build it? The lab cites curiosity, better understanding, and safety — Block’s example: imagine being in serious pain while a voice AI has no idea and just keeps going.
How Creators and Freelancers Can Profit While You Wait
Honest framing first: no API, no pricing, no public access. The money angle is preparation, not a launch:
- Sell voice UX audits. Test a client’s voice bot for interruptions, emotion, and pauses. Dessa just reset what clients expect from “natural” voice.
- Build a voice demo reel. Record natural conversation flows with today’s tools, so you have scripts and proof when speech-native models open up.
- Offer voice persona packages. Scripts, backchannel patterns, and interruption handling transfer directly to speech-native models.
- Pitch privacy-first voice. If local laptop inference becomes real, clinics, law offices, and shops that cannot send audio to the cloud are a ready market.
- Create content on the shift. Explainers comparing speech-native with text-pipeline voice are fresh and timed to a news moment.
Winners of platform shifts rarely wait for the launch email — they already have clients and demos.
What Dessa Does Not Prove Yet
- No benchmarks. Recordings, not test scores — outsiders cannot compare Dessa with any rival.
- Selected demos. Block chose the recordings; it does not say how many conversations went badly.
- A hedged “first” claim. Block says “as far as we know” Dessa is the first realistic live conversation model trained only from human speech.
- Untested local claims. Nobody outside Block has confirmed the MacBook or faster-than-real-time claims.
- Real safety questions. A voice this realistic raises impersonation and scam risks — hence safety-researcher-only early access, per Block.
Frequently Asked Questions
What is Block Dessa?
A full-duplex speech-to-speech model from Block’s small research lab, trained from scratch on conversation only, per Block.
Can I try Dessa today?
No. No public access, API, weights, or pricing announced — early access is for safety researchers only.
How is Dessa different from Siri or ChatGPT voice?
Today’s assistants turn speech to text, think in text, and read a reply aloud. Dessa, per Block, never touches text: audio in, audio out.
What hardware does Dessa need?
Block says a MacBook Pro M4 Max or better, running locally. That claim has not been independently tested.
Who built Dessa?
Block’s small AI research lab, founded by former members of Toronto startup Dessa, reporting to Jack Dorsey and led by Ragavan Thurairatnam.
When will Dessa be publicly available?
No date announced. Watch block.xyz/dessa — and treat anything else claiming to be Dessa as unofficial.
Keep Reading
Sources
- Block’s official Dessa page: https://block.xyz/dessa — source of all claims here, labelled as Block’s.
No YouTube video embedded: none found from Block’s official channel.