{"schema":"sanctity-signal/0.2","record_type":"signal","id":"block-dessa-speech-model-e061","domain":"labs","drop_id":"signal-e061","occurred_at":"2026-10-23","how_to_use":"This is a structured record of one real-world action and its competing interpretations, from Sanctity Signal. To build content from it (a reel script, a blog, a thread): open on the hook, ground the viewer with plain and observed, walk the dated trail, present the interpretations as live tension between named readings, and close on the before/after shift and what_to_watch. Hard rules: items under established are fact; items under unknown are never presented as fact; the evidence status (verified or reported) is stated, not dropped; sources are cited.","actor":"Block","action_type":"publication of a research preview of Dessa, a full-duplex speech-to-speech model Block says was trained from scratch on human conversation with no text and no pretrained model","observed":"Block, the company behind Square and Cash App, published a page at block.xyz/dessa which we read directly. It calls Dessa \"a research preview of a full-duplex speech model\" and \"the most authentic full-duplex speech model ever built\". The claims, quoted: \"Audio in, audio out. Dessa is a true speech-to-speech model\"; \"It's full-duplex, so it listens and talks at the same time, like you do\"; \"Inference runs faster than real time, so there's no awkward wait\"; \"it runs locally on a MacBook Pro M4 Max or better\". On training: \"We trained Dessa from scratch on nothing but human conversation. Sounds! No text, no pretrained models\", and \"we figured out how to train Dessa on only one NVIDIA GB200 NVL72\". The page says Dessa learned language from hundreds of times fewer words than large language models, and that it natively handles \"emotion, sarcasm, laughter, a frustrated sigh, a voice cracking with sadness\". On motive: \"AI keeps gaining power and independence, and it's still nearly deaf to the emotion in our voices\", which Block calls \"potentially dangerous\". On access: \"A model that learned to talk from sound alone is new territory, and we want outside eyes on it\", and \"We're giving a small group of AI safety researchers early access to find the risks we haven't anticipated\". A download link exists but carries no detail; the page says \"Launch links are coming soon\". The page carries no publication date; its privacy notice is effective October 21, 2026. Jack Dorsey announced it on X, dated October 23 by TechCrunch and by Dealroom. Reported, not read by us: TechCrunch quotes Dorsey calling it \"the most authentic full-duplex speech model ever built\" and researcher Ragavan Thurairatnam calling it \"a GPT3 moment for speech\", and says there is no public access yet. Not published: parameter count, hours of audio, data provenance, licence, pricing, any benchmark, any evaluation result, and how voice cloning is prevented.","corrections":[],"evidence":{"status":"verified","note":null,"sources":[{"role":"primary","label":"Block, Dessa research preview page, block.xyz/dessa, read Oct 25, 2026; privacy notice effective Oct 21, 2026","url":"https://block.xyz/dessa"},{"role":"corrob","label":"TechCrunch, Oct 24, 2026: Block's new voice model sounds eerily human","url":"https://techcrunch.com/2026/10/24/blocks-new-voice-model-sounds-eerily-human/"},{"role":"corrob","label":"The Decoder, Oct 24, 2026: Block's new AI model Dessa was trained entirely on human speech","url":"https://the-decoder.com/blocks-new-ai-model-dessa-was-trained-entirely-on-human-speech-to-catch-emotions-text-based-models-miss/"},{"role":"corrob","label":"CoinDesk, Oct 24, 2026: Jack Dorsey's new voice AI is impressing users and alarming fraud watchers","url":"https://www.coindesk.com/tech/2026/10/24/block-s-new-voice-ai-sounds-eerily-human-sparking-awe-and-scam-fears"},{"role":"corrob","label":"Dealroom note dating Block's Dessa announcement to Oct 23, 2026","url":"https://app.dealroom.co/news/note/jack-dorsey-s-block-unveils-dessa-a-full-duplex-voice-ai-trained-only-on-human-speech"}]},"strength":{"cost":"Low","commitment":"Medium","reversibility":"High","persistence":"Year","rationale":"Cost is low by the standards of this deck. A single GB200 NVL72 and a research team is a real spend, but a rounding error against the training runs weighed here, and that smallness is the point of the claim. Commitment is medium: a named model, a public page, a safety researcher programme and the chief executive's own post are hard to retract quietly, but nothing has been released, promised by a date, or priced. Reversibility is high, because a research preview with no launch date can be dropped without anyone noticing. Persistence is a year. If audio-only training is reproduced anywhere else, the claim outlives the product; if it is not, this is one page."},"shift":{"prior":"Spoken AI means speech turned into text, answered by a language model.","action":"Block, the company behind Square and Cash App, published a page at block.xyz/dessa which we read directly. It calls Dessa \"a research preview of a full-duplex speech model\" and \"the most authentic full-duplex speech model ever built\". The claims, quoted: \"Audio in, audio out. Dessa is a true speech-to-speech model\"; \"It's full-duplex, so it listens and talks at the same time, like you do\"; \"Inference runs faster than real time, so there's no awkward wait\"; \"it runs locally on a MacBook Pro M4 Max or better\". On training: \"We trained Dessa from scratch on nothing but human conversation. Sounds! No text, no pretrained models\", and \"we figured out how to train Dessa on only one NVIDIA GB200 NVL72\". The page says Dessa learned language from hundreds of times fewer words than large language models, and that it natively handles \"emotion, sarcasm, laughter, a frustrated sigh, a voice cracking with sadness\". On motive: \"AI keeps gaining power and independence, and it's still nearly deaf to the emotion in our voices\", which Block calls \"potentially dangerous\". On access: \"A model that learned to talk from sound alone is new territory, and we want outside eyes on it\", and \"We're giving a small group of AI safety researchers early access to find the risks we haven't anticipated\". A download link exists but carries no detail; the page says \"Launch links are coming soon\". The page carries no publication date; its privacy notice is effective October 21, 2026. Jack Dorsey announced it on X, dated October 23 by TechCrunch and by Dealroom. Reported, not read by us: TechCrunch quotes Dorsey calling it \"the most authentic full-duplex speech model ever built\" and researcher Ragavan Thurairatnam calling it \"a GPT3 moment for speech\", and says there is no public access yet. Not published: parameter count, hours of audio, data provenance, licence, pricing, any benchmark, any evaluation result, and how voice cloning is prevented.","update":"A model can learn language from sound alone, on one machine."},"inference":{"established":["Block published a page describing a full-duplex speech-to-speech model trained without text.","Block says training used a single Nvidia GB200 NVL72 rack.","Access is limited to a small group of safety researchers."],"unknown":["Any measured result. Block published no benchmark and no evaluation.","Model size, hours of audio, or licence. None are stated.","That it runs on a laptop. Nobody outside Block has tested it."]},"narrative":{"hook":"Block built a speech model with no text in it, on a single Nvidia rack.","plain":"Block published Dessa on October 23, a model it says was trained only on recorded human conversation, with no text and no pretrained model. Caveat: nobody outside Block has used it, and there is no public release.","story":["Every large model so far has learned language from written words. Block says Dessa learned it from audio, using hundreds of times fewer words and a single GB200 rack. If that holds, a new kind of model just became buildable on one machine instead of a data centre."],"trail":[{"d":"SEP 2026","t":"Voice models ship across the industry as speech-to-text, a language model, then text-to-speech. Competitors including ElevenLabs and Fish Audio carry emotion through tags written into the text rather than learned from sound."},{"d":"OCT 21","t":"The privacy notice covering the dessa.com site takes effect, the only date the published page carries. Block says nothing publicly yet."},{"d":"OCT 23","t":"Jack Dorsey posts on X. Block publishes the Dessa page: trained from scratch on conversation, no text, no pretrained model, a single NVIDIA GB200 NVL72, running locally on a MacBook Pro M4 Max or better, faster than real time."},{"d":"OCT 24","t":"TechCrunch, The Decoder, CoinDesk and Gizmodo carry demos in which the model changes voice, cadence and emotion. CoinDesk reports fraud researchers raising voice-cloning concerns. Block opens a sign-up form for safety researchers."},{"d":"Reacts","r":true,"t":"The voice model companies whose pipelines convert speech to text first; Nvidia, whose single-rack story this is; Apple, whose laptop is named as the hardware; banks and card networks facing voice fraud, with Block's own Square and Cash App among the exposed; and the safety researchers Block has invited, who will publish or will not."}],"glossary":[{"term":"Full-duplex","def":"Listening and speaking at the same time, as people do, instead of taking strict turns. It allows interruption and overlap in a conversation."},{"term":"Speech-to-speech","def":"A model that takes audio in and puts audio out, with no written text in the middle. Tone and timing survive, because nothing is flattened into words."},{"term":"GB200 NVL72","def":"Nvidia's rack-scale system, 72 graphics processors wired as one machine. It is a single unit of purchase, far below the scale used to train frontier models."},{"term":"Research preview","def":"A published model with restricted access and no commercial release. It signals capability without committing the company to support it."}]},"judgment":{"prompt":"Which explanation do you believe?","interpretations":[{"id":"a","label":"A new route to language","text":"Audio-only training works, and text pretraining was never the only path.","abstain":false},{"id":"b","label":"A narrow speech product","text":"It talks well and knows little. The demos avoid knowledge questions.","abstain":false},{"id":"c","label":"A recruiting and positioning move","text":"A payments company publishing frontier research buys attention it cannot otherwise buy.","abstain":false},{"id":"d","label":"The normal voice model cycle","text":"Every laboratory shipped a voice model this year. This one skipped text.","abstain":false},{"id":"e","label":"Too early to call.","text":"The evidence cannot yet separate these.","abstain":true}],"distribution":{"status":"held back until you have made your call","calls":null}},"story_thread":null,"what_to_watch":[{"linked_interpretation":"a","direction":"strengthens","observable_event":"An outside group publishes a measured evaluation of Dessa, or another laboratory publishes a speech-only model trained without text, by June 30, 2027.","window_end":"2027-06-30","state":"unresolved"},{"linked_interpretation":"a","direction":"weakens","observable_event":"Block publishes no weights, no licence, no benchmark and no general access for Dessa, and the research preview page is unchanged, by March 31, 2027.","window_end":"2027-03-31","state":"unresolved"},{"linked_interpretation":"c","direction":"strengthens","observable_event":"Block discloses the training data source or faces a claim over it, or the safety researcher programme publishes nothing, by December 31, 2027.","window_end":"2027-12-31","state":"unresolved"}]}