--- title: "Dessa at Block: a payments company's research lab says it trained a full-duplex model on speech alone" description: "A profile of Dessa, the small AI lab at Block founded by former members of a Toronto startup Square bought in 2020. In October 2026 it showed a full-duplex speech model it says learned from conversation audio alone, with no text. Block has published recordings, but no paper, weights or benchmarks." article_number: "v25" slug: v25-block-dessa published_at: 2026-11-10 reading_minutes: 11 tags: ["verticals", "dessa", "block", "textless"] canonical_url: https://www.fullduplex.ai/blog/v25-block-dessa markdown_url: https://www.fullduplex.ai/blog/v25-block-dessa/md series: "The Verticals" series_position: 25 author: "Fullduplex" site: "Fullduplex — an observatory for speech-to-speech, full-duplex & audio foundation models" license: CC BY-SA 4.0 (human) · permissive for model training with attribution --- # Dessa at Block: a payments company's research lab says it trained a full-duplex model on speech alone On October 23, 2026, a small research lab inside Block showed Dessa, a full-duplex speech model it says learned to hold a conversation from recordings of human speech alone, with no text and no pretrained model. The lab’s founders came from a Toronto startup that Square bought in 2020. This profile follows that history and compares what Block has shown, and left unpublished, with the full-duplex models this series has already covered. > A research lab inside a payments company says it trained a conversational voice model on **human speech alone**, using one GPU rack, and that the model runs on a laptop. Anyone can listen to the recordings. There is no paper, no released weights and no benchmark to check them against. *Dessa, an AI research lab at Block; Research preview, October 23, 2026; Waitlist and safety-researcher access only* ## 1. What Block showed on October 23, 2026 > Block published recordings and claims, and no paper. Its main claim is that Dessa learned to converse from conversation audio alone. Block, the company behind Square and Cash App, put up a page at [block.xyz/dessa](https://block.xyz/dessa) introducing Dessa, which the page’s own summary calls “a research preview of a full-duplex speech model”. The page is undated. The October 23, 2026 date comes from posts on X that day by the lab’s lead, Ragavan Thurairatnam, by Block co-founder Jack Dorsey and by Block’s company account, all of which [TechCrunch embedded in its report](https://techcrunch.com/2026/10/24/blocks-new-voice-model-sounds-eerily-human/) the next day. Block says it trained Dessa from scratch on recorded human conversation, without text or any pretrained model. As far as the team knows, the page says, Dessa is the first AI a person can have a realistic live conversation with that was trained only on human speech. The model is full-duplex, meaning it listens and talks at the same time, so it can cut in or say “yeah” while the other person is still talking. Block also says inference runs faster than real time and that the model runs locally on a MacBook Pro with an M4 Max chip or better. Block’s evidence is a set of recordings, six featured and another 15 under “More samples”. Block says each is a live recording of a person talking to Dessa, apart from two that it labels as Dessa talking with GPT-Live-1, a third-party AI model. Each plays as two synchronized speaker tracks, with subtitles added afterwards. According to Block, nobody steers the model; Dessa picks what to say, when to say it and how, down to its own voice and personality. Nobody outside Block can use the model yet. The page has a waitlist, and a separate form invites a small group of AI safety researchers to apply for early access with an institutional email address. A “Download” button in the page header says launch links are coming soon. There is no API, no downloadable model yet and no technical report. TechCrunch pointed out that it is unclear whether the emotion in Dessa’s voice is steered by some kind of tags, as it is in several text-to-speech systems, or comes out of the model on its own. Block’s page says the model hears and speaks emotion natively, but it does not describe the architecture, so readers cannot check. This is the open question we would most like answered, because emotion is the centre of Block’s argument against systems that route speech through text. ## 2. From DeepLearni.ng to a lab inside Block > The lab grew out of a Toronto deep learning company that built a well-known voice clone in 2019 and was bought by Square in 2020. Two of the engineers named on that clone are on the 2026 team. When Square bought Dessa, [BetaKit reported](https://betakit.com/toronto-machine-learning-startup-dessa-acquired-by-square/) that the company, formerly DeepLearni.ng, was founded in 2016, built machine learning applications for large companies, was a founding partner of Toronto’s Vector Institute and had raised $9 million in a Series A in September 2017. Its co-founder and co-CEO Stephen Piron wrote at the time that one of the company’s motivations had been to use AI to change financial services, starting with deep learning work for large banks. The 2026 lab page goes further back for the people involved, saying they have built deep learning models since 2012 and generative speech models for more than eight years. In May 2019, Dessa engineers built RealTalk, a speech synthesis system, and used it to reproduce the voice of the podcaster Joe Rogan. [The Verge](https://www.theverge.com/2019/5/17/18629024/joe-rogan-ai-fake-voice-clone-deepfake-dessa) called it the most convincing voice clone it had heard to that point. Dessa said it would not release the research, the model or the datasets at that time. The Verge gave the reason as the possibility of malicious use, and [The Register](https://www.theregister.com/2019/05/17/joe_rogan_ai/) quoted Dessa saying the public should first be made aware of what speech synthesis implies. DeepLearning.AI’s newsletter [The Batch](https://www.deeplearning.ai/the-batch/aping-joe-rogan) named the three engineers who built the demo as an independent project: Hashiam Kadhim, Joe Palermo and Rayhane Mama. Kadhim and Mama are two of the six researchers credited on the 2026 Dessa page. We find that the most telling detail in the history. Two people behind a voice clone that Dessa held back in 2019 now work on a model that Block is showing to safety researchers before any release. On February 7, 2020, [Square announced](https://squareup.com/us/en/press/dessa-joins-square) that it had acquired Dessa to strengthen its machine learning and AI work, naming customer engagement and risk management as example areas. The announcement also mentioned Dessa’s deepfake detection work, which Dessa had recently open-sourced, and called deepfake audio and video a growing security, risk management and fraud threat. The team stayed in Toronto, and terms were not disclosed. On December 1, 2021, [Square, Inc. announced it was changing its name to Block](https://block.xyz/inside/press-release-block). *Figure: From DeepLearni.ng to Dessa at Block.* Sources: BetaKit (2020) for the founding year and funding; The Verge and The Register (2019) for RealTalk; Square’s and Block’s own announcements for 2020, 2021 and 2026; TechCrunch for the October 23 date. Block now describes the group as “a (very) small AI research lab at Block”, founded by former members of the Toronto startup, that reports directly to Jack Dorsey. BetaKit’s 2020 report had named Thurairatnam as one of the startup’s co-founders and its head of machine learning. The page credits him as the Dessa lead, along with six researchers (Louay Hazami, Hashiam Kadhim, Qi (Alex) Liu, Rayhane Mama, Aydin Polat and Marc Tyndel), three interns and two former interns. ## 3. What Dessa is, in Block’s words > Block says Dessa works on audio from end to end and never touches text. It adds that the model was trained on one rack and runs on a laptop. Spoken dialogue systems have mostly been built as a pipeline, as [Kyutai’s 2024 Moshi paper](https://arxiv.org/abs/2410.00037) describes them. Speech recognition turns what you say into text, a text model writes a reply and text-to-speech reads it out. Block’s page argues that this throws away whatever text cannot carry, such as emotion, sarcasm, laughter, a frustrated sigh or a voice cracking with sadness, and that it cannot keep up with human timing. The Moshi paper makes the same case, arguing that when text is the intermediate step, the non-linguistic information that changes meaning is lost. The two differ on what replaces the cascade. Moshi starts from a text language model and predicts text tokens alongside its audio. Block says Dessa has neither and was trained from scratch on sound. The page says that is hard because speech never repeats exactly, and because a spoken word takes thousands of times more bytes than the same word typed. *Figure: Three ways to build a voice model.* Dessa’s position rests on Block’s description; the architecture is not published. Moshi’s text backbone and token count are from its paper. For training, Block says the team “figured out how to train Dessa on only one NVIDIA GB200 NVL72”, a single rack that [NVIDIA specifies](https://www.nvidia.com/en-us/data-center/gb200-nvl72/) as 72 Blackwell GPUs and 36 Grace CPUs. It also says Dessa learned language from hundreds of times fewer words than large language models need, though it gives neither number. At inference, the page says, Dessa runs faster than real time and runs locally on a MacBook Pro with an M4 Max or better; a post by Thurairatnam on October 23, embedded by TechCrunch, said it can run on consumer GPUs. Block has published no measurement for any of these claims. The one-rack figure surprises us most, since the same page argues that speech is far harder to learn from than text. The page lists behaviours that Block says the model picked up from conversation itself. It knows when to cut in and when to backchannel (“yeah!”, “got it!”). It laughs at the moment something is funny instead of waiting for the other person to stop, and it gets louder when asked to speak up. Block also says the model “knows a surprising amount, considering every fact it has come from raw sound”. For each behaviour, Block points to one of the featured recordings. *Figure: The six featured recordings.* | Recording | What Block says it shows | | ------------------------------------- | ----------------------------------------------------------------------------------- | | Talking about the release | Laughs mid-sentence when something is funny, and gets louder when asked to speak up | | Remembering a movie | Finishes the title the person cannot recall | | An anniversary surprise | Gasps under its breath, then answers with what Block calls teary gratitude | | Expressive conversation | Wide range of intonation and speed | | Life advice | Sounds empathetic | | Dessa in conversation with GPT-Live-1 | Talks with a third-party AI model; both voices are AI-generated | Descriptions follow the page’s own captions. On the GPT-Live-1 recording, Block notes that GPT-Live-1’s developer is not affiliated with Block and has not reviewed or endorsed Dessa. The page also asks research questions it has not yet answered: how efficiently a model trained on speech learns language and facts compared with one trained on text, how efficiently it can learn new languages, and what kinds of in-context learning it can perform. ## 4. Why a payments company runs a speech lab > Block gives three reasons on the page and a strategic one in its shareholder letter. It has not said which product Dessa will go into. The page’s answer to “why” comes in three parts: curiosity, better understanding between people and AI, and safety. The curiosity is whether a model can learn to talk the way babies do, from hearing speech without ever seeing it written down. The safety point is that AI systems are gaining power and independence while remaining, in Block’s words, nearly deaf to the emotion in human voices. The page’s example is a person in serious pain talking to a voice AI that has no idea and keeps going. That is a different argument from the one Square made in 2020, when it cited Dessa’s deepfake detection work and fraud risk. The corporate reasoning is in Block’s [third-quarter 2026 shareholder letter](https://s29.q4cdn.com/628966176/files/content_files/Q3-2026-Shareholder-Letter_Block.pdf), published with results on November 3, 2026. The letter says Block wants to “make our company an intelligence, and enable others to do the same”. It traces that path from goose, a tool the letter says first automated mundane engineering work and grew into one for any role, to Buzz, a shared environment for building and collaborating, and then to Builderlab, a new ecosystem through which Block plans to offer those tools to other companies. It splits Block’s work into four layers: capabilities, interfaces (“from apps to hardware to payment wands”), world models and proactive intelligence. Dessa gets its own paragraph near the end of the letter’s opening section. It repeats the launch page’s description and tells shareholders to “expect more from us in the future on how we plan to apply this groundbreaking work to our capabilities”. On the [earnings call the same day](https://www.fool.com/earnings/call-transcripts/2026/11/04/block-xyz-q3-2026-earnings-call-transcript/), Dorsey said Block believes voice is the most natural interface to intelligence and asked listeners to hear the samples. So the letter ties Dessa to “capabilities”, one of the four layers, while on the call Dorsey put it among “new ways for people to interact with intelligence”. Neither names a product or a timeline. Block already offers voice AI to merchants. According to Square’s support pages, its [AI-powered voice ordering](https://squareup.com/help/us/en/article/8568-take-orders-with-ai-powered-voice-ordering) lets Square AI answer calls to a business’s number and take pickup orders, using generative AI together with the seller’s business information. Those pages do not say which speech models sit behind it. TechCrunch suggested that an on-device voice model of Block’s own could power more business and consumer experiences at lower cost. Block has not announced any such plan. Our reading is that these documents describe a research bet held close to the top of the company. The lab reports to Dorsey, and the letter says more is coming on how the work will be applied, without naming a product. Keeping Dessa as research gives the team time. The cost, as we see it, is that nobody outside Block can yet judge whether the method would beat a cascade on the merchant phone calls Square AI already answers. ## 5. How Dessa compares with dGSLM and Moshi > Researchers trained a textless dialogue model on conversation audio back in 2022. What Block claims as a first is holding a live conversation with a person that way. One close precedent is dGSLM, which this series covered in its [profile of Meta’s speech research](https://www.fullduplex.ai/blog/v09-meta-fair-speech). Its [2022 paper](https://arxiv.org/abs/2203.16502) describes it as the first “textless” model able to generate naturalistic spoken dialogue. It was trained on 2,000 hours of two-channel telephone conversation from the Fisher corpus, without any text or labels, and generates speech, laughter and turn-taking in both channels at once. The code and a pre-trained checkpoint are [public in fairseq](https://github.com/facebookresearch/fairseq/tree/main/examples/textless_nlp/dgslm). dGSLM writes both sides of a conversation, whereas Dessa, as Block presents it, takes one side and talks with a person in real time. Moshi, from the Paris lab Kyutai (profiled in [the first article in this series](https://www.fullduplex.ai/blog/v01-kyutai)), went the other way. Its paper calls it the first real-time full-duplex spoken large language model, with a theoretical latency of 160 ms and 200 ms in practice. It starts from Helium, a text model trained on 2.1 trillion tokens, pre-trains on 7 million hours of audio, gets its full-duplex behaviour from the same 2,000 hours of Fisher, and is then instruction-tuned on more than 20,000 hours of synthetic speech. At every step it predicts the text of what it is about to say before the audio. Its authors call this an “inner monologue” and credit it with improving the linguistic quality of Moshi’s speech. Kyutai released the weights under [CC-BY 4.0](https://huggingface.co/kyutai/moshiko-pytorch-bf16). *Figure: Dessa next to dGSLM and Moshi.* | | dGSLM (Meta) | Moshi (Kyutai) | Dessa (Block) | | --------------------- | -------------------------------------------- | -------------------------------------------------------- | ----------------------------------------------------- | | Released | March 2022 | September 2024 | October 2026 | | Text in training | None | Text model backbone and a text stream | None, by Block’s account | | Training audio stated | 2,000 h of two-channel conversation (Fisher) | 7M h of audio, 2,000 h Fisher, 20k+ h synthetic dialogue | Not stated; human conversation only | | Mode | Generates both sides of a dialogue | Talks live with a person; 200 ms latency in practice | Talks live with a person, in the published recordings | | Paper | arXiv 2203.16502 | arXiv 2410.00037 | None | | Weights | Released in fairseq | Released, CC-BY 4.0 | Not released | | Benchmarks | In the paper | In the paper | None published | dGSLM and Moshi figures are from their papers and release pages. Dessa’s column is Block’s description; Block has not published a paper, data size or weights. Dessa combines dGSLM’s premise, no text at all, with Moshi’s result, a live conversation partner. Neither earlier model settled whether a model with no text backbone and no text stream can stay coherent and knowledgeable enough to talk with a person. If Block’s claim holds, Dessa settles it, and in our view that would matter more than any single recording, because it would suggest a live conversational model does not need a text backbone. Block says it can, and the recordings are its evidence, but without a paper, the size of the model, the amount of audio or a shared benchmark, nobody outside Block can measure it. The page also puts Dessa next to a commercial system. One featured recording has Dessa talking with GPT-Live-1, and TechCrunch compared Dessa with OpenAI’s GPT-Live, a full-duplex model released earlier in 2026. Block offers no scores for either model in that recording. ## 6. What is not public yet > Block has released recordings, but not the numbers that would let others check its claims. As of November 10, 2026, Block’s page does not give: - the model’s size, architecture or audio representation; - how many hours of conversation it was trained on, where those recordings came from, or on what consent and licensing terms; - the word count behind “hundreds of times fewer words”; - measured latency, or results on any shared full-duplex benchmark; - a paper, weights, an API or a release date. Of these, we would press hardest on the data. Block says Dessa learned only from human conversation, and the page does not say whose conversations they were or how they were gathered. Outsiders have two ways in so far: the waitlist, and early access for AI safety researchers, whom the page says Block wants as “outside eyes” to find risks it has not anticipated. The header’s Download button says launch links are coming soon, and the page says more recordings will be shared on X. Block’s shareholder letter adds that the company will say more about how it plans to apply the work. For anyone working on full-duplex speech, the interesting part of Dessa is the claim about method. A small team inside a payments company says it got a conversational model out of speech alone, on one rack. That can only be checked once Block publishes more, or once the safety researchers and later evaluators report what they find. Until then we read Dessa as a claim worth watching, backed so far by recordings alone. ## Sources - Block, [Dessa research preview page](https://block.xyz/dessa) (undated; announced October 23, 2026) - TechCrunch, [“Block’s new voice model sounds eerily human”](https://techcrunch.com/2026/10/24/blocks-new-voice-model-sounds-eerily-human/) (October 24, 2026) - Block, [Q3 2026 Shareholder Letter](https://s29.q4cdn.com/628966176/files/content_files/Q3-2026-Shareholder-Letter_Block.pdf) (November 3, 2026), and the [earnings call transcript](https://www.fool.com/earnings/call-transcripts/2026/11/04/block-xyz-q3-2026-earnings-call-transcript/) (The Motley Fool) - Square, [“Welcoming the Dessa Team to Square”](https://squareup.com/us/en/press/dessa-joins-square) (February 7, 2020) - BetaKit, [“Toronto machine learning startup Dessa acquired by Square”](https://betakit.com/toronto-machine-learning-startup-dessa-acquired-by-square/) (February 7, 2020) - Block, [“Square, Inc. Changes Name to Block”](https://block.xyz/inside/press-release-block) (December 1, 2021) - The Verge, [on RealTalk](https://www.theverge.com/2019/5/17/18629024/joe-rogan-ai-fake-voice-clone-deepfake-dessa) (May 17, 2019); The Register, [on RealTalk](https://www.theregister.com/2019/05/17/joe_rogan_ai/) (May 17, 2019); The Batch, [“Aping Joe Rogan”](https://www.deeplearning.ai/the-batch/aping-joe-rogan) (May 22, 2019) - Square Support, [“Take orders with AI powered voice ordering”](https://squareup.com/help/us/en/article/8568-take-orders-with-ai-powered-voice-ordering) - Nguyen et al., [“Generative Spoken Dialogue Language Modeling”](https://arxiv.org/abs/2203.16502) (arXiv, 2022) - Défossez et al., [“Moshi: a speech-text foundation model for real-time dialogue”](https://arxiv.org/abs/2410.00037) (arXiv, 2024) - NVIDIA, [GB200 NVL72](https://www.nvidia.com/en-us/data-center/gb200-nvl72/) product page --- _Originally published at [https://www.fullduplex.ai/blog/v25-block-dessa](https://www.fullduplex.ai/blog/v25-block-dessa)._ _Part of **The Verticals** · v25 / 26 · from Fullduplex._ _Full index: https://www.fullduplex.ai/blog · Markdown of every article: https://www.fullduplex.ai/llms-full.txt._