Jack Dorsey’s financial firm Block has introduced the Dessa AI model, an artificial intelligence system that learns to hold spoken conversations without ever reading text or relying on written training data [1]. Instead of converting vocal audio into written transcripts before formulating an answer, Dessa processes and generates speech through a direct sound pipeline that mirrors how children pick up words long before they ever learn how to read [2]. Dessa learned without written words [1].
Speech Training Without Text Data
Block trained Dessa from scratch using recorded human dialogue, avoiding the written text datasets and pretrained foundations that power most voice assistants found on smartphones and home speakers. Diego Almada Lopez at Crypto Briefing reported that Dessa learned spoken language while hearing hundreds of times fewer words than conventional text models absorb during development [2]. The engineering team wanted to test whether an artificial intelligence system could pick up spoken dialogue in the same way infants learn to talk without ever seeing printed words [1].
According to project details shared by Block, human infants pick up conversational skills without ever seeing a single printed word, and testing whether software could follow that same developmental trajectory proved challenging but technically possible through direct audio modeling [1]. Block carried out this training on a single NVIDIA GB200 NVL72 system, showing that speech-first artificial intelligence can train on specialized hardware clusters without depending on massive server farms [2]. The development team stated that speech serves as the most natural interface to any intelligence, but conventional voice setups flatten human conversation into plain text and struggle with natural timing. When voice software ignores vocal rhythm, AI systems remain deaf to the emotion behind a person’s words [1].
The project was created inside Block by a research team led by Ragavan Thurairatnam, whose members previously worked at the Toronto startup Dessa before Dorsey acquired the company. Thurairatnam directly reports to Dorsey. These researchers have spent several years developing generative speech tools for Block. Block hasn’t disclosed model weights or detailed performance benchmarks, choosing instead to share select audio recordings showing how Dessa navigates real-time dialogue [2].
Full Duplex Audio Processing
Dessa relies on a full duplex architecture that lets the system listen and speak at the same time, avoiding the sequential pauses common to cloud voice assistants. Standard assistants wait until a person finishes talking before sending an audio recording to an outside server for transcription. Because Dessa processes raw sound continuously, it can register conversational pauses, manage interruptions, and adjust vocal delivery without resetting the dialogue [2].
By processing audio directly instead of reading transcribed sentences, Dessa identifies laughter, sarcastic phrasing, and pitch shifts that text parsers discard [2]. The developers argued that current voice assistants remain deaf to human feeling, creating safety problems when someone seeks help during a sudden emergency. If a person speaks while suffering from physical pain, a typical assistant might treat the words as a routine text prompt and keep talking. “That’s the safety problem,” Block said [1].
Block outlined three reasons for building Dessa: curiosity (aka fun), better understanding between people and artificial intelligence, and safety [1]. Instead of relying on canned replies, Dessa shapes its pitch and vocal speed to match the human speaker. The model treats voice as sound, not text [2]. Aman Shukla at Skoar noted that this approach contrasts with mainstream assistants designed for rigid utility, pointing toward dialogue systems that feel closer to an awkward friend than a corporate search tool [3].
How the Dessa AI Model Responded in Tests
To display what the software can do, Block shared several audio recordings of the Dessa AI model chatting with humans, revealing exchanges that can be remarkably fluid in places but also awkward and strange as the system navigates casual conversation. In one demonstration shared by project lead Ragavan Thurairatnam, Dessa carried on a steady back-and-forth exchange with a speaker, though it comically mistook the person for someone named “Ronaldo” at the start. Dessa comically mistook him for someone named Ronaldo [1]. Another recorded sample showed Dessa playing a voice-based guessing game, where it asked questions, parsed spoken clues from the human player, and reacted to hints without derailing the round [3].
Other audio recordings featured conversations about cinema, including an exchange about movies directed by Christopher Nolan, where Dessa unexpectedly veered into an unprompted tangent about Superman films directed by Bryan Singer. In a separate interaction, a speaker told Dessa about plans for a surprise vacation to Paris, prompting an emotional vocal response. The voice answered with teary gratitude. It thanked the speaker like an eager travel partner [1].
Block also recorded a direct audio conversation between Dessa and OpenAI’s GPT-Live-1 voice model, pairing the two systems together to contrast how a speech-first model talks when compared to a voice tool rooted in text processing. In that exchange, Dessa came across as noticeably more casual and natural in its spoken phrasing than the competing synthetic voice from OpenAI. OpenAI built its model around text foundations. Jay Peters reported for The Verge that while these recordings show unusual conversational flow, Block selected and published the examples itself, and independent evaluators have not verified Dessa in unscripted environments [1].
Local Hardware Requirements
Block engineered the Dessa AI model to run inference locally on personal computer hardware instead of sending private voice streams across the web. According to company specifications, Dessa operates on an Apple MacBook Pro equipped with an M4 Max chip or better. Dessa requires an Apple M4 Max chip. Running speech-to-speech models locally eliminates latency caused by internet round trips, allowing Dessa to respond to interruptions without waiting for cloud servers [2].
Local processing keeps sensitive voice recordings on the user’s machine, addressing privacy concerns that arise when proprietary audio is stored on external corporate servers. The engineering team hasn’t announced whether standard Apple laptops or Windows computers will eventually run Dessa, leaving the M4 Max as the only documented baseline. For now, local execution demonstrates that compact speech systems can function without relying on server farms [2].
The company hasn’t disclosed how much system memory Dessa consumes during active conversations or whether Block will share its code with open-source developers [2]. Coverage from Discernion noted that running generative audio on consumer hardware represents progress toward decentralized artificial intelligence, even though training such models from scratch remains technically difficult [4]. Block hasn’t set a commercial release date for wider consumer access [2].
Evaluating the Dessa AI Model for Safety
Instead of distributing Dessa to the general public, Block plans to grant early access to a select group of AI safety researchers [1]. These independent specialists will examine how an artificial intelligence model trained purely on human speech behaves across long conversations, looking for potential risks before Block considers wider distribution. Safety experts will study how Dessa handles sensitive topics and whether its conversational quirks lead to misleading statements [2].
The development team noted that voice interfaces require cautious study because speech creates a perception of empathy that written text cannot match. When an artificial voice laughs, hesitates, or sounds touched by a story, listeners can easily mistake the algorithm for a caring human friend [1]. A report from AIToolly observed that Block’s focus on awkward, emotional speech dynamics contrasts with mainstream tools built purely for efficiency, raising fresh safety questions about emotional attachment to artificial voices [5].
For now, Dessa remains an experimental research initiative inside Jack Dorsey’s company rather than a commercial payment feature [2]. The research behind the Dessa AI model proves that artificial intelligence can pick up conversational speech directly from sound, bypassing written text just as human infants do [1]. Whether those awkward, emotional interactions can turn into dependable everyday software will depend on findings from upcoming safety reviews [2].
- ONLINE NEWS Peters, J. (2026, October 23). Jack Dorsey’s Block made an AI model that talks to you like an awkward new friend. The Verge. [Article Link]
- ONLINE NEWS Lopez, D. A. (2026, October 23). Jack Dorsey’s Block unveils Dessa, an AI voice model trained entirely on human conversations. Crypto Briefing. [Article Link]
- ONLINE NEWS Shukla, A. (2026, October 23). Block Unveils AI Model That Chats Like an Awkward New Friend. Skoar. [Article Link]
- WEBSITE Discernion. (2026, October 23). Jack Dorsey’s Block Makes an AI Model That Talks Like an Awkward New Friend. Discernion. [Article Link]
- ONLINE NEWS AIToolly. (2026, October 24). Jack Dorsey’s Block Announces Dessa AI Model Trained Exclusively on Human Conversation. AIToolly. [Article Link]
- ONLINE NEWS La Primera en Línea. (2026, October 23). Jack Dorsey’s Block made an AI model that talks to you like an awkward new friend. La Primera en Línea. [Article Link]
