Dessa Voice AI: What It Is, Availability, Features and Reviews
Last verified: October 28, 2026 (Asia/Shanghai). Facts below come from Block's official Dessa page, Jack Dorsey's and Ragavan Thurairatnam's posts on X, TechCrunch, The Verge, CoinDesk, and the Hacker News thread linked in Sources. Block has published no model card, parameter count, price, API docs or independent benchmarks. We have not tested Dessa hands-on; nobody outside Block has published a hands-on review yet. Demo clips are vendor-selected recordings.
Dessa Voice AI is Block's full-duplex speech-to-speech conversation model, announced on October 23, 2026. On the official page, Block calls it "the most authentic full-duplex speech model ever built," trained "from scratch on nothing but human conversation" with "no text, no pretrained models." Jack Dorsey posted the same framing on X the same day (announcement). Lead Ragavan Thurairatnam called it "a GPT3 moment for speech" (post). Below: what Dessa is, what is (and is not) available, how early listeners reacted, who it is for, and how it stacks up against a shipped rival such as Grok Voice 3.0.
| At a glance | Dessa Voice AI |
|---|---|
| Maker | Small AI research lab at Block (Square / Cash App); reports to Jack Dorsey |
| What it is | Full-duplex speech-to-speech conversation model trained only on human speech audio |
| Status | Research preview (Oct 23–28, 2026); early access for a small group of AI safety researchers |
| Public product / API / price | None published |
| Local run (vendor claim) | MacBook Pro M4 Max or better |
| Training hardware (vendor claim) | One NVIDIA GB200 NVL72 |
| Official benchmarks | None published |
| Contact | ragavan+dessa@block.xyz (listed on the official page) |
Sources: block.xyz/dessa, TechCrunch, The Verge, CoinDesk.
Practical rule: treat Dessa as a research demo you can listen to, not a product you can buy, download or call via API—until Block publishes access, pricing and docs.
What Dessa Voice AI is
Block's lab describes Dessa as true audio-in / audio-out speech-to-speech. It is full-duplex: it "listens and talks at the same time," can cut in, and can backchannel with short acknowledgements such as "yeah!", "got it!" and "ooh" (official page). The same page says inference "runs faster than real time," and that the model "natively hears and speaks emotion, sarcasm, laughter, a frustrated sigh, a voice cracking with sadness."
That is the core pitch against cascaded voice stacks (speech-to-text → LLM → text-to-speech). Block argues those pipelines "turn your voice into text and throw away an incredible amount of information text alone can't carry." Knowledge, on Block's telling, comes from conversational audio rather than a separate text LLM—TechCrunch summarizes that "Dessa's knowledge stems from the conversational data it has" (TechCrunch).
The lab is "founded by former members of Dessa, the Toronto deep learning startup," which Square acquired; the team says it has built generative speech models for over eight years and reports directly to Dorsey (official page; historical context in CoinDesk). Researchers named on the page include Louay Hazami, Hashiam Kadhim, Qi (Alex) Liu, Rayhane Mama, Aydin Polat and Marc Tyndel; Ragavan Thurairatnam is listed as Dessa lead.
Demo clips on the site are labeled as live person-to-Dessa recordings with subtitles added afterward. Examples include mid-sentence laughter, finishing a forgotten movie title, a surprise-trip reaction described as "teary gratitude," and a conversation with OpenAI's GPT-Live-1 (Block notes GPT-Live-1's developer is not affiliated and has not endorsed Dessa).
Preview frame from X/Twitter CDN; may be slow or blocked in some regions.
Availability, pricing and API
As of October 28, 2026:
- No public download, waitlist open to consumers, app store listing, API, SDK or pricing page appears on block.xyz/dessa.
- Block is "giving a small group of AI safety researchers early access" and asks interested researchers to get in touch (official page; echoed by TechCrunch and The Verge).
- CoinDesk notes Block "hasn't disclosed model size, benchmarks or a release date" (CoinDesk).
- Vendor hardware claims only: local inference on a MacBook Pro M4 Max or better; training on one NVIDIA GB200 NVL72 (official page). Jack Dorsey separately replied on X that he saw "zero latency on my macbook" when quoting a local-run comment (post)—that is a personal claim, not a published latency benchmark.
Watch for lookalikes. Third-party sites that offer a "Dessa" voice chat, TTS credits or a download today are not Block's research preview. Block's only official product page for this model is block.xyz/dessa.
Practical rule: if a site sells "Dessa API access" or a one-click install in October 2026, it is not the Block research preview described here.
Specs Block has (and has not) published
| Claim | Source | Notes |
|---|---|---|
| Speech-to-speech, audio in / audio out | Official page | Vendor description |
| Full-duplex (listen + speak together) | Same | Vendor description |
| Trained only on human conversation audio; no text; no pretrained models | Same | Vendor claim; not independently audited |
| Faster-than-real-time inference | Same | Vendor claim; no published ms figures |
| Runs locally on MacBook Pro M4 Max or better | Same | Vendor claim |
| Trained on one NVIDIA GB200 NVL72 | Same | Vendor claim |
| Parameter count / architecture / languages / context length | — | Not published |
| Official benchmarks / model card / paper | — | Not published |
| Public API / pricing | — | Not published |
Block frames three motivations: "curiosity (aka fun), better understanding between humans and AI, and safety," including the worry that today's voice AI is "nearly deaf to the emotion in our voices" (official page).
Official benchmarks
None. Block has not published Speech-to-Speech Index scores, Full Duplex Bench numbers, latency percentiles, or a model card. Any chart you see online that assigns Dessa a numeric leaderboard score without citing Block or a named third-party lab is unverified—do not trust it. For comparison context only, shipped models such as Grok Voice 3.0 do publish vendor tables (still vendor-reported unless an independent source repeats them).
Real-world reactions (demos only)
Nobody outside Block has reported a hands-on session. Reactions are about Block's curated audio/video demos.
Official / founder posts (verbatim):
- Jack Dorsey: "We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence." (X)
- Ragavan Thurairatnam: "We made the first AI model you can have a realistic conversation with that was trained entirely on speech! Dessa is a full-duplex speech-to-speech model that you can run on consumer GPUs. This is a GPT3 moment for speech." (X)
Press: TechCrunch titled its piece "Block's new voice model sounds eerily human" and noted Square already uses voice AI for ordering, while Dessa could "power a lot more business and consumer experiences at a lower cost" if it stays on-device (TechCrunch). The Verge called demos occasionally "a little awkward and uncanny," including a clip where Dessa initially mistook a speaker for "Ronaldo" (The Verge). CoinDesk reported both awe and scam fears around human-sounding phone AI (CoinDesk).
Hacker News (thread, ~533 points / ~199 comments at fetch time) was split:
- toddmorey: "The conversation examples are impressive but this made me say holy shit out loud: 'it runs locally on a MacBook Pro M4 Max or better. And we figured out how to train Dessa on only one NVIDIA GB200 NVL72.'" (comment)
- weego: "Sorry I'm lost. The sample on the homepage sounds absolutely awful… it's so far off passing as 'real' that I feel like I must be missing something." (comment)
- IshKebab: "Have you tried the alternatives? This is clearly far far superior in terms of fluidity of conversation." (comment)
- bushido: "It is kinda scary that we're here already, I can see this being pretty problematic for seniors/public prone to getting scammed." (comment)
- haarts: "Ok. Sounds cool. Now what? Open source weights? An API somewhere?" (comment)
Third-party press image from TechCrunch; may be slow or blocked in some regions.
Watch on YouTube: "Dessa by Jack Dorsey's Block" by Fahd Mirza
Third-party video by Fahd Mirza, an independent creator. AI Takeaway did not make this video and is not affiliated with its creator or with Block.
No time to watch? AI Takeaway is a tool that works on any YouTube video, including third-party explainers like this one. Paste the link to get key points in minutes, jump to original timestamps, and ask follow-ups with answers that cite the video. Browse more AI and tech interviews in the Hub.
What launch coverage and reviews add
Use video and social demos critically: timestamp-verifiable claims (what the audio literally does) beat marketing slogans ("most authentic," "GPT3 moment"). Cross-check every product claim against block.xyz/dessa and note when a clip is vendor-selected. Safety commentary in CoinDesk and on HN is about possible misuse (phone scams, deceptive agents); Block itself is inviting safety researchers precisely because speech-only full-duplex models are "new territory."
Who should care about Dessa Voice AI
- AI safety and speech researchers: the only named early-access audience on the official page.
- Voice-product builders watching the frontier: useful as a reference for full-duplex, emotion-preserving speech-to-speech—even if you keep shipping cascaded or API-hosted stacks today.
- On-device / edge teams: the M4 Max-class local claim is the headline hardware story; verify yourself when weights or binaries exist.
- Probably not for production owners this week: no API, SLA, languages list, tool-calling story or price. If you need a callable realtime voice agent now, look at shipped options such as Grok Voice 3.0 instead.
How it compares (briefly)
| Criterion | Dessa Voice AI (Block) | Grok Voice 3.0 (xAI / SpaceXSI) |
|---|---|---|
| Maturity | Research preview, demos only | Shipped API model (grok-voice-3.0) |
| Training story | Speech-only, no text (vendor) | Speech-to-speech agent model with published docs |
| Access | Safety-researcher early access | Public realtime WebSocket API |
| Price | Not published | $0.08/min (vendor) |
| Local laptop claim | M4 Max or better (vendor) | Hosted API (see docs) |
| Published benchmarks | None | Vendor tables vs rivals (see our Grok Voice 3.0 guide) |
Dessa is interesting because of how it was trained and how human the demos sound; Grok Voice 3.0 is interesting because you can pin a model ID and pay per minute today. For broader model context beyond voice, see Gemini 4 Argon, GPT-6.1 Sol vs GPT-6 Astra, Muse Wonder, DeepSeek V4.1-Flash-1017 and DeepMind's Genie 4.
Practical rule: pick Dessa to study the speech-only research bet; pick a documented API voice model when you have customers on the phone this quarter.
Getting-started checklist
- Listen to the official demos on block.xyz/dessa—not third-party "Dessa" chat sites.
- Read Jack Dorsey's and Ragavan Thurairatnam's launch posts for the vendor framing (jack, Ragavan).
- If you are an AI safety researcher, email the contact on the official page; otherwise wait for a public waitlist, download or API.
- Skim the HN thread for skeptical takes on intelligence vs. fluency.
- If you need production voice agents now, evaluate a shipped API (for example Grok Voice 3.0) with your own call audio.
- Re-check block.xyz/dessa before assuming open weights, pricing or language support—none of those were published as of this verification date.
Keep up without the watch time
Voice and frontier launches are moving weekly. If you would rather read than watch long interviews, the AI Takeaway video summary hub has free curated summaries with timestamp jumps, for example:
- Muse Wonder: Zuckerberg and Wang's $100B AI
- Gavin Baker on Orbital Compute, TSMC, and Frontier Models
- Sam Altman on AGI, Compute, and Human Agency
Install the AI Takeaway Chrome extension for key points on YouTube, or submit any YouTube video on aitakeaway.net. More guides live on the articles index; keep checking the Hub as new coverage lands.
FAQ
What is Dessa Voice AI?
Dessa is Block's full-duplex speech-to-speech conversation model, announced October 23, 2026. Block says it was trained from scratch on human conversation audio only—no text and no pretrained models—and that talking with it feels like talking with a person (official page).
Can I download or call Dessa today?
Not publicly. As of October 28, 2026, Block offers early access only to a small group of AI safety researchers and has not published a consumer download, API or price (official page; TechCrunch).
Does Dessa run offline on a laptop?
Block claims local inference on a MacBook Pro M4 Max or better (official page). Independent confirmation is not available yet because the model is not publicly distributed.
Is there a Dessa API or pricing?
No official API or pricing has been published. Ignore third-party sites that sell "Dessa API" access under that name unless they clearly document a different product.
How does Dessa compare to Grok Voice 3.0?
Dessa is a speech-only research preview with striking demos and no public API. Grok Voice 3.0 is a shipped realtime API voice model with published pricing and docs—see our Grok Voice 3.0 guide.
Are the demo conversations real?
Block states every clip on the page is a live recording of a person talking to Dessa, with nobody steering responses and subtitles added afterward (official page). They are still vendor-selected examples, not an independent eval set.
Sources
- Block, Dessa – the Conversation Model trained only on speech (official product/research page)
- Jack Dorsey on X, Dessa announcement and local MacBook reply (Oct 23, 2026)
- Ragavan Thurairatnam on X, "GPT3 moment for speech" and Meet Dessa (Oct 23, 2026)
- TechCrunch, Block's new voice model sounds eerily human (Oct 24, 2026)
- The Verge, Jack Dorsey's Block made an AI model that talks to you like an awkward new friend (Oct 23, 2026)
- CoinDesk, Jack Dorsey's new voice AI is impressing users and alarming fraud watchers (Oct 24, 2026)
- Hacker News, Dessa: Conversation Model trained only on speech (and comments linked above)
- YouTube (third-party): Fahd Mirza – Dessa by Jack Dorsey's Block
- Related AI Takeaway guides: Grok Voice 3.0, Kimi K3.1, Muse Wonder, Gemini 4 Argon, GPT-6.1 Sol vs GPT-6 Astra, DeepSeek V4.1-Flash-1017