--- title: Block unveils Dessa, a full-duplex speech model trained... description: Block unveiled Dessa, a full-duplex speech-to-speech model trained only on human voice conversations, with no text data and no pretrained models involved. Jack... canonical: https://daily.dev/posts/block-unveils-dessa-a-full-duplex-speech-model-trained-only-on-voice-wglco6kby twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Block unveils Dessa, a full-duplex speech model trained only on voice | daily.dev og:description: Block unveiled Dessa, a full-duplex speech-to-speech model trained only on human voice conversations, with no text data and no pretrained models involved. Jack... og:url: https://daily.dev/posts/block-unveils-dessa-a-full-duplex-speech-model-trained-only-on-voice-wglco6kby og:image: https://api.daily.dev/og/posts/wGLCO6KBY.png og:image:alt: Block unveils Dessa, a full-duplex speech model trained only on voice og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Block unveils Dessa, a full-duplex speech model trained only on voice **[Collections](https://daily.dev/sources/collections)** · 2 min read · 0 upvotes · 0 comments ## Summary Block unveiled Dessa, a full-duplex speech-to-speech model trained only on human voice conversations, with no text data and no pretrained models involved. Jack Dorsey announced it, calling voice the most natural way to interact with intelligence. The model is small enough to run on a MacBook, and access is now open to AI safety researchers. Only demo samples are public so far; @testingcatalog says the clips go well beyond anything else in emotional expression and natural feel, though that's one reaction to a handful of samples. Separately, the SEC's fiscal 2027 exam priorities add AI to its watch list, with examiners stepping up scrutiny of AI risks and private fund valuations. ## Content Block, Jack Dorsey's company, has unveiled Dessa, a speech-to-speech AI model. Dorsey's pitch: voice is the most natural way to interact with intelligence, and Dessa is meant to be the most authentic full-duplex speech model yet. ## What Block says it is - Dessa was trained on conversational audio only. There is no text data and no pretrained model underneath it. - It is full-duplex, so it can listen and talk at the same time. That lets it handle pauses and interruptions, much like OpenAI's GPT-Live. - Emotion is built into the model. Block contrasts this with systems like ElevenLabs v4, which steer emotion with voice tags. Demo clips include laughs, pauses and cries. - The company says it runs locally on a MacBook Pro M4 Max or better. One account says it was trained on a single NVIDIA GPU. ## Availability Only demo samples are out. There is no public access yet, and Block is asking AI safety researchers to test the model first. ## My take The demos are what people are reacting to. @testingcatalog says they go well beyond anything else in emotional expression and natural feel. I'd hold off on that verdict until more people can try it, but the claim that a model this expressive runs on a laptop is the part I find most interesting. ## Questions this post answers ### What is Dessa, the speech model from Block? Dessa is a full-duplex speech-to-speech model from Block, announced by Jack Dorsey, trained solely on human voice conversations with no text data and no pretrained models used. It is small enough to run on a MacBook, and access is now open to AI safety researchers, though only demo samples have been released publicly so far. _Anyone tracking new approaches to voice AI can follow releases like Dessa on daily.dev._ ### Does Dessa use text data or pretrained language models for training? No, Dessa uses no text data and no pretrained models at all. It is trained solely on human voice conversations, making it a speech-to-speech model built entirely from audio, which is unusual compared to most voice AI systems that build on top of text-based language models. _Developers comparing voice model training approaches can track details like this on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 5 discussions and 73 comments across x (as of 2026-10-25). **TL;DR:** Reactions are split between fascination with a voice-only, full-duplex speech model and heavy skepticism about whether it can handle facts, tool use, or real-world noisy conversations beyond polished demo clips. **Sentiment:** 30% positive · 40% mixed · 30% skeptical **The case for** - Several found the natural pauses, backchannels, and emotional cadence impressive compared to typical text-first voice bots. - Some see promise for call centers, phone support, and on-device/private voice assistants running locally on a MacBook. - A few liked that skipping text pretraining could make interfaces feel less robotic and more natural. **The pushback** - Many questioned how a model trained with zero text data can learn facts or world knowledge. - Commenters doubted it can reliably call tools or pass structured data (like reading a URL or arguments) without any text representation. - Some pointed out full-duplex models without an LLM backbone typically see quality drop compared to cascaded or Thinker-Talker architectures. - Skeptics said demos are easy and want to see performance on interruptions, bad mics, or emotionally charged live conversations. - Some reactions were dismissive or mocking, questioning the voice choice, usefulness, or calling it scary/unnecessary. **By community** - x (mixed): Reactions mix genuine technical curiosity about voice-only training and full-duplex architecture with persistent doubts about factual grounding and tool use, plus a separate thread with more dismissive, mocking, or unsettled reactions to the voice itself. **Hottest debate:** Whether a model trained purely on voice with no text data can actually retain facts and perform tasks, or whether it's just cadence and emotion without substance. **Open questions** - How does Dessa handle factual questions or world knowledge without any text pretraining? - Can it call tools or pass structured data like URLs or arguments without a text layer? - How does it perform with interruptions, accents, or noisy/bad-mic real-world conditions versus polished demos? - Does quality hold up in live, emotionally charged conversations rather than scripted demo clips? **Highlights** > @lzuren2001 @testingcatalog On fullduplex without LLMs the quality of the conversations usually drops sharply. That’s why Thinker-Talker (that uses an LLM as a “brain”) architectures are so popular. More then that, quality usually drops with any new modality: with duplex it’s worse than with an LLM… > — [korovsky on x · 1 comments](https://x.com/korovsky/status/2113877067125424274) > @testingcatalog trained only on voice conversations with no text data is wild. how does it handle factual questions without any text pretraining? > — [yevloop on x · 1 points](https://x.com/yevloop/status/2113895584008388911) > @testingcatalog full duplex is the part i'm most curious about. listening while talking makes "wait, stop" mid sentence easy. the harder case is stop mid action. in my tests a cascaded agent told a caller "the refund was canceled" and the refund landed ~350 ms later. no tools in dessa yet, would > — [ChrisKennLazar on x](https://x.com/ChrisKennLazar/status/2113776260602290270) > @testingcatalog Skipping text entirely is the interesting bet. Every mainstream voice model treats speech as text wearing a costume — tokens in, tokens out. If semantics can emerge from raw conversational audio alone, the text-first pipeline stops being obviously correct. > — [Alex20060101 on x](https://x.com/Alex20060101/status/2113807046038741187) > @testingcatalog Demo samples always sound emotional. The real test is a live conversation where someone's upset and the model has to respond instead of perform. Did any of the demos show that? > — [Soulsopic on x](https://x.com/Soulsopic/status/2113780838122807573) **Source threads** - [x](https://x.com/feross/status/2113878394270302262) · 0 points · 0 comments - [x](https://x.com/kloss_xyz/status/2113771852288123147) · 0 points · 0 comments - [x](https://x.com/testingcatalog/status/2113766526545232108) · 1 points · 46 comments - [x](https://x.com/willsmithesssss/status/2113709069210923229) · 0 points · 0 comments - [x](https://x.com/TheInsiderPaper/status/2113763618538057983) · 384 points · 27 comments ## Similar posts on daily.dev - [OpenAI's GPT-Live: ChatGPT voice that listens and talks](https://daily.dev/posts/openai-s-gpt-live-chatgpt-voice-that-listens-and-talks-nwap9djkb) · The Next Web · 9 upvotes · 3 comments --- Tags: [#voice-ai](https://daily.dev/tags/voice-ai) [View this post on daily.dev](https://daily.dev/posts/block-unveils-dessa-a-full-duplex-speech-model-trained-only-on-voice-wglco6kby) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Block unveils Dessa, a full-duplex speech model trained only on voice","url":"https://daily.dev/posts/block-unveils-dessa-a-full-duplex-speech-model-trained-only-on-voice-wglco6kby","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/block-unveils-dessa-a-full-duplex-speech-model-trained-only-on-voice-wglco6kby"},"datePublished":"2026-10-24T06:21:20.087Z","dateModified":"2026-10-25T01:19:10.012Z","description":"Block unveiled Dessa, a full-duplex speech-to-speech model trained only on human voice conversations, with no text data and no pretrained models involved. Jack...","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/block-unveils-dessa-a-full-duplex-speech-model-trained-only-on-voice-wglco6kby","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"voice-ai","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Block unveils Dessa, a full-duplex speech model trained only on voice"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/block-unveils-dessa-a-full-duplex-speech-model-trained-only-on-voice-wglco6kby#faq","mainEntity":[{"@type":"Question","name":"What is Dessa, the speech model from Block?","acceptedAnswer":{"@type":"Answer","text":"Dessa is a full-duplex speech-to-speech model from Block, announced by Jack Dorsey, trained solely on human voice conversations with no text data and no pretrained models used. It is small enough to run on a MacBook, and access is now open to AI safety researchers, though only demo samples have been released publicly so far. Anyone tracking new approaches to voice AI can follow releases like Dessa on daily.dev."}},{"@type":"Question","name":"Does Dessa use text data or pretrained language models for training?","acceptedAnswer":{"@type":"Answer","text":"No, Dessa uses no text data and no pretrained models at all. It is trained solely on human voice conversations, making it a speech-to-speech model built entirely from audio, which is unusual compared to most voice AI systems that build on top of text-based language models. Developers comparing voice model training approaches can track details like this on daily.dev."}}]} ```