Block has unveiled Dessa, a speech-to-speech model that the company says it trained from scratch on nothing but human conversation, with no text and no pretrained models. The model listens and speaks at the same time, which is what engineers mean by full-duplex. For anyone who has talked over a voice assistant and watched it freeze or restart, that design goal is the whole story.
AI Chat Daily and Digg both date the announcement to October 23, 2026, and both describe Dessa as a research preview rather than a product. Block has not announced pricing, a general release date, or broad access. Early access goes to a small group of AI safety researchers, and a public waitlist is open, according to those reports.
This analysis walks through what Block has claimed, which details come only from secondary reporting, and what is still missing. It also places Dessa next to the wider voice AI field and the on-device push we covered in our look at the best local LLMs of 2026.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Block Dessa at a Glance: What Was Announced
The short version fits in a list. Everything below comes from Block’s project page or from outlets that read it, and the later sections flag which items rest on secondary reporting only.
- Company: Block, the fintech company behind Square and Cash App.
- Model: Dessa, a full-duplex, speech-to-speech model.
- Training approach: from scratch, using only human conversation, with no text and no pretrained models.
- Behavior: listens and speaks at once, interrupts, and offers short backchannels such as “yeah!”, “got it,” and “ooh.”
- Status: research preview, with recorded demos, a waitlist, and early access for a small group of AI safety researchers.
- Date: October 23, 2026, per AI Chat Daily and Digg.
Block frames the release as a new category rather than a faster version of an old one. Its page says Dessa opens a different class of models with new behaviors for researchers to study. The company hedges its headline claim, too. It says that as far as it knows, Dessa is the first AI that supports a realistic live conversation after training on human speech alone.
That hedge matters. A “first” claim in AI is easy to make and hard to audit, and Block has published no benchmark that would let outsiders test it. We treat the claim as Block’s position, not as an established fact.
What a Full-Duplex Speech Model Actually Does
Most voice assistants work in turns. You talk, the system detects silence, converts your speech to text, generates a text reply, and reads it aloud. A full-duplex speech model drops the turn-taking. It processes incoming audio while it produces outgoing audio, the way two people overlap on a phone call.
Audio in, audio out
AI Chat Daily reports that Dessa takes audio in and puts audio out, with no transcription step in between. Block’s page calls it a true speech-to-speech model and says everything it knows came from human speech. That detail separates Dessa from voice products built around a text language model with a speech layer bolted on each end.
The practical consequence is information. Text strips out tone, pacing, hesitation, and laughter. A model that never converts to text can, in principle, keep all of it. Block says Dessa hears and speaks emotion, sarcasm, laughter, a frustrated sigh, and a voice cracking with sadness.
Interruptions and backchannels
Two behaviors define natural conversation, and Block highlights both. The first is cutting in at the right moment. The second is the backchannel, a quick “yeah” or “got it” that tells the speaker you are following without taking the floor. Block says Dessa learned both from real conversations, not from hand-written rules.
Block’s page also describes demo clips in which Dessa laughs mid-sentence at something funny and speaks louder when asked. The company says nobody steers the model in those recordings. Treat that as a vendor description of vendor-selected clips, since no independent tester has published results yet.
Trained From Scratch on Human Speech: The Core Claim
The training story is the part of the announcement that most separates Dessa from existing voice systems. Block says it trained the model from scratch on human conversation alone. No text corpus and no pretrained language model sit underneath it. In Block’s words, it learned to hold a real-time conversation simply by listening.
Most voice products today borrow their knowledge from a text model trained on a vast pile of written material. Dessa claims a different path. Every fact it knows reached it as sound. Block’s page makes this point itself, noting that the model knows a surprising amount considering all of it came from raw audio.
AI Chat Daily adds one efficiency claim. Block says Dessa learned language from “hundreds of times fewer words” than large language models. The announcement gives no precise word count, so readers cannot check the ratio. It is a headline-grabbing comparison with no denominator.
Block frames the larger payoff as research. Its page asks how efficiently speech-trained models learn language and facts compared with text, how efficiently they learn new languages, and what kinds of in-context learning they can do. Those are open questions, and Block is inviting researchers to probe them.
One GB200 NVL72 to Train, a MacBook Pro to Run
Two hardware claims travel with the announcement. According to PANews, training used only a single NVIDIA GB200 NVL72 system, and the finished model runs locally on a MacBook Pro with an M4 Max chip or better at faster-than-real-time speed. AI Chat Daily reports the same two details.
Both details come from secondary reports, so we attribute them to those outlets. If they hold, they change the cost story. Frontier voice and language models usually need clusters. A single rack-scale system for training, and a laptop for inference, point to a much smaller footprint.
Digg’s roundup also surfaces a social post attributed to the project lead, Ragavan Thurairatnam, saying the model can run on consumer GPUs. That is a second local-hardware route beyond the MacBook, though it appears in a post and not in a spec sheet. Anyone planning around it should wait for published requirements.
Local execution matters for voice more than for text. Round trips to a data center add delay, and delay is exactly what breaks overlapping speech. It also keeps raw audio on the device, a privacy angle that resonates with the hardware trend we tracked in AMD’s Ryzen AI Halo, which targets 300B-class local models.
Claims, Sources, and Verification Status
Not every detail in circulation rests on the same footing. The table below sorts the main claims by who reported them and how firmly each is established. “Official page” means Block’s project page. “Secondary” means outlets summarizing it.
| Claim | Reported by | Status |
|---|---|---|
| Dessa is a full-duplex speech-to-speech model | Block’s page, AI Chat Daily, Digg | Stated by Block, consistent across outlets |
| Trained from scratch on human conversation only | Block’s page, Digg, PANews | Stated by Block, not independently audited |
| No text and no pretrained models | Block’s page, AI Chat Daily | Stated by Block, not independently audited |
| Announced October 23, 2026 | AI Chat Daily, Digg, The Verge timestamp | Secondary reports, not an official-page fact |
| Trained on one NVIDIA GB200 NVL72 | AI Chat Daily, PANews | Secondary reports |
| Runs locally on MacBook Pro M4 Max or better | Block’s page, AI Chat Daily, PANews | Stated by Block, no independent test yet |
| Early access for AI safety researchers | AI Chat Daily, Digg, PANews | Secondary reports |
| Learned from “hundreds of times fewer words” | AI Chat Daily | Block claim, no word count published |
The pattern is clear. The architecture and training claims come straight from Block and agree across outlets. The numbers that make for easy headlines, such as the hardware and the data-efficiency ratio, are the ones hardest to verify today.
What Block Has Not Disclosed Yet
AI Chat Daily states plainly that Block has not published a parameter count, a dataset size, latency figures, or formal benchmark results. Reports also show no price and no general release date. For a technical audience, those gaps define how much weight the announcement can carry right now.
| Detail | Published? | What reports say |
|---|---|---|
| Parameter count | No | AI Chat Daily says none published |
| Training dataset size | No | AI Chat Daily says none published |
| Latency figures | No | AI Chat Daily says none published |
| Formal benchmark results | No | AI Chat Daily says none published |
| Price | No | No pricing in any source reviewed |
| General release date | No | Broader access remains a waitlist, per Digg |
| Public weights or code | Not confirmed | No source confirms an open release |
| Early access program | Yes | Small group of AI safety researchers |
One caution on terminology. Some social replies summarized by Digg call Dessa an open-sourced model. None of the primary or news sources we reviewed confirms released weights or code. Until Block says so, treat Dessa as a restricted research preview.
Why Text-Mediated Voice AI Falls Short
Block’s argument starts with a complaint. Its page says current voice AI turns your voice into text and throws away a huge amount of information that text cannot carry. The company says this leads to a worse experience and is potentially dangerous, because even the most powerful AI systems cannot hear how a person feels.
AI Chat Daily summarizes the same logic. Most assistants convert speech to text before reasoning about it, which strips emotional cues. The table frames the contrast as Block presents it. It is a design comparison, not a measured head-to-head.
| Dimension | Typical text-mediated voice AI (as Block describes it) | Dessa (as Block describes it) |
|---|---|---|
| Input handling | Speech converted to text first | Audio in, audio out, no transcription step |
| Knowledge source | Text-trained language model | Human conversation audio only |
| Turn-taking | Waits for the speaker to finish | Listens and talks at the same time |
| Interruptions | Often scripted or detected by silence | Learned from real conversation |
| Backchannels | Rare | “yeah!”, “got it!”, “ooh” |
| Emotional cues | Lost in transcription | Heard and expressed natively |
| Where it runs | Usually cloud | Locally on MacBook Pro M4 Max or better |
Block does not claim that text models are bad at reasoning. Its claim is narrower and easier to defend. If your goal is a natural spoken exchange, routing everything through text discards signal you cannot get back.
What Block Said in Its Own Words
Three lines from Block’s project page capture the pitch. We quote them exactly as published. On training, the page states: “We trained Dessa from scratch on nothing but human conversation.” The next sentence is just as blunt: “No text, no pretrained models.”
On how the model behaves, Block’s page says: “It’s full-duplex, so it listens and talks at the same time, like you do.” The wording is plain on purpose. Block wants readers to picture a conversation, not a pipeline diagram.
Block’s headline claim comes with its hedge attached: “As far as we know, Dessa is the world’s first AI you can have a realistic live conversation with that was trained only from human speech.” The introduction on the page is shorter: “Meet Dessa, our speech model trained from scratch on human conversation.”
Digg’s summary says Dessa’s project lead, Ragavan Thurairatnam, called the release a GPT-3 moment for speech. We paraphrase that framing because it reaches us through a social post, not a document we could verify. It signals how Block wants the release read, as a step-change for a modality and not a feature update.
Early Reactions From The Verge and Social Media
The first press read was mixed in tone. The Verge’s Jay Peters wrote that Block’s shared chats with humans can occasionally be a little awkward and uncanny, a phrase Digg’s roundup also quotes. His headline compared Dessa to an awkward new friend, which is a fair summary of an early demo with no polish pass.
Social reaction ran warmer. Digg’s AI summary reports 68.2 percent positive and 31.8 percent negative sentiment, based on only 23 sentiment-bearing replies from 22 accounts in one conversation. That sample is tiny, so read it as a mood check and not a poll. Supporters praised emotional expressiveness and the prospect of private local hosting, while skeptics called the project scammy or dismissed the results as unremarkable.
One more detail from Block’s page deserves attention. One demo shows Dessa talking with a third-party model labeled GPT-Live-1. Block says that model’s developer is not affiliated with Block and has not reviewed or endorsed Dessa. Both voices in that clip are AI-generated. Block is plainly comfortable inviting comparison with an established voice product.
Who Else Is Building Voice Models
Dessa enters a crowded field. Our research names OpenAI’s real-time voice experience, Kyutai’s Moshi, Sesame’s conversational voice systems, and Google’s Gemini Live as the closest reference points. We could not verify a fresh version, release, or capability update for each of those within the past one to three months, so we make no head-to-head capability claims.
What we can say is where Block tries to differentiate. The company is not selling lower latency alone. It is selling a model whose conversational behavior was learned from speech alone, without text supervision or a pretrained language model. That is a statement about origins, and origins are hard for rivals to copy after the fact.
The bigger platforms are also moving voice into agents. Our comparison of ChatGPT Dots, Meta Muse, and Claude agents shows how quickly assistant products are being repackaged around autonomous work. A speech-native model is a different bet, aimed at the quality of the conversation itself.
The practical difference for developers is dependency. Products built on a text model inherit its pricing, rate limits, and update schedule. A speech-native model that runs on a laptop would let a small team ship a voice feature with no per-minute cloud bill. Block has not said whether or when outside builders can use Dessa, so that remains a possibility, not a roadmap.
Historical Context: From Command-and-Response to Conversation
Voice assistants spent years as command interpreters. They recognized a phrase, mapped it to an action, and spoke a canned answer. Large language models upgraded the middle of that chain, and the wave of voice modes that followed made replies smarter without changing the turn-based structure.
The remaining friction was always rhythm. People overlap, trail off, laugh, and say “mm-hm” while the other person talks. A pipeline built on speech recognition, a text model, and speech synthesis has to guess when you are done, and that guess is where the awkward pauses and clipped interruptions come from.
Full-duplex research targets that gap directly. Instead of detecting the end of a turn, the model keeps listening while it speaks and decides moment by moment whether to continue, yield, or interject. Dessa’s twist is the training diet. Block bets that learning from real two-sided conversation, with nothing else mixed in, teaches those timing skills better than adding them to a text model afterward.
Whether that bet pays off is an empirical question. The history of AI is full of cases where a bigger or broader training mix beat a purer one. Block’s own page frames Dessa as a way to find out, and the early-access program exists to generate that evidence.
Market Impact for Fintech, Developers, and Local AI
Block is a payments company, and a voice model might look like a detour. PANews describes Dessa as the work of the AI research lab inside Block, and AI Chat Daily says that lab reports directly to a Block cofounder and chairman. Neither outlet says Dessa will ship inside a Block product, and we will not guess at one.
The link to finance is easy to imagine anyway. Customer support, fraud calls, and account recovery are voice-heavy workflows where tone and hesitation carry meaning. At the same time, banks are wary of agents they do not control, as we covered when five big US banks moved to block third-party AI agents. A speech model that talks like a person raises the stakes on identity and consent in those settings.
For the wider market, the signal is that a non-traditional AI entrant can claim frontier-style novelty with a small training footprint. That puts pressure on incumbents whose voice products sit on top of text models. It also arrives while the assistant market is still reshuffling, as the figures in our Claude, ChatGPT, Gemini, and Grok usage share breakdown show.
For hardware buyers, the local angle is the news. If Dessa-class models run on a MacBook Pro M4 Max, high-memory laptops gain a new reason to exist beyond coding and image work. Apple-silicon machines and consumer GPUs become inference targets for conversational AI, not just text chat.
Safety and Access: Why Researchers Go First
Block chose a cautious rollout. According to The Verge’s reporting as relayed by Digg, the company is inviting a small group of AI safety researchers to request early access and look for risks its own team has not anticipated. AI Chat Daily adds that applicants must apply with a name and an institutional email from a university, company, or lab.
The caution fits the technology. A model that hears emotion and responds with laughter, sighs, and a cracking voice can persuade in ways a flat text reply cannot. It can also imitate human warmth convincingly, which opens obvious doors to manipulation and impersonation. Block’s own page calls text-only voice AI potentially dangerous, yet a more human-sounding model brings its own risks.
Researchers will likely probe a short list of questions. Can the model be steered into deceptive intimacy? How does it behave under distress or crisis conversations? What does it retain from the audio it hears? Block has not published answers, and the waitlist suggests broader testing will follow only after this first phase.
Five Predictions for the Next Six Months
These are our reads, not reported facts. We base them on what Block has disclosed and on how earlier AI research previews tended to evolve. Readers tracking the broader field can follow our AI and machine learning coverage for updates.
- Block publishes numbers. Researchers will ask for parameter count, latency, and benchmark data, and the early-access group gives Block a reason to release at least some of it.
- Rivals showcase overlap and interruption. Expect competing voice teams to lead with turn-taking and backchannel demos, since Dessa made those behaviors the benchmark for “natural.”
- Independent tests land first on the claims people can run. Local speed on Apple silicon is the easiest claim to check, so expect early hands-on reports once access widens.
- The “first” claim faces challenges. Other labs will argue that earlier full-duplex or speech-only systems already exist, and Block’s hedge, “as far as we know,” leaves room for that debate.
- Access widens slowly, with guardrails. The waitlist converts to a limited release before any open-ended launch, given the impersonation risks of an emotionally fluent voice.
The wildcard is whether Block ties Dessa to a product. Nothing in the reports says it will, and a pure research posture is consistent with everything published so far.
FAQ: Block Dessa Questions Answered
What is Block’s Dessa? Dessa is a full-duplex, speech-to-speech AI model from Block. The company says it trained the model from scratch on human conversation alone, with no text and no pretrained models.
When was Dessa announced? AI Chat Daily and Digg date the announcement to October 23, 2026. Block’s own page does not clearly state the date, so treat the date as secondary-source reporting.
What does full-duplex mean? It means the model listens and speaks at the same time instead of waiting for the other person to finish. That lets it interrupt, overlap, and give quick acknowledgments such as “yeah” or “got it.”
Can I use Dessa today? Not broadly. Reports describe a research preview with recorded demos, a public waitlist, and early access for a small group of AI safety researchers. Block has announced no price and no general release date.
What hardware does Dessa need? According to AI Chat Daily and PANews, it runs locally on a MacBook Pro with an M4 Max chip or better, and training reportedly used a single NVIDIA GB200 NVL72 system. A social post relayed by Digg also mentions consumer GPUs. Block has not published a full requirements sheet.
Is Dessa open source? No source we reviewed confirms released weights or code. Some social replies describe it as open-sourced, but Block describes a research preview with restricted access.
How does Dessa differ from other voice assistants? Block says most voice AI converts speech to text first and loses emotional cues. Dessa takes audio in and puts audio out with no transcription step, and its knowledge came only from human speech.
Has Block published benchmarks for Dessa? No. AI Chat Daily reports that Block has not published a parameter count, dataset size, latency figures, or formal benchmark results.
