Dessa

This article is about the 2026 speech model from Block. For the Toronto machine learning company of the same name that Square acquired in 2020, see the Dessa (company) section below.

Dessa is a full-duplex, speech-to-speech conversational AI model developed by a small research lab at Block, Inc., the company behind Square and Cash App.^[8] Block announced it on 23 October 2026 as a research preview. The company says Dessa was trained from scratch on recordings of human conversation alone, "no text, no pretrained models", and that it listens and speaks at the same time, produces audio faster than real time, and runs locally on a MacBook Pro with an M4 Max chip or better.^[1] Block cofounder Jack Dorsey promoted the release on X, calling Dessa "the most authentic full-duplex speech model ever built".^[2] As of 30 October 2026 the model was not publicly available: Block was offering early access to a small group of AI safety researchers, and the download and waitlist buttons on its page were placeholders marked as coming soon.^[1][8]

The lab is led by Ragavan Thurairatnam and was founded by former members of Dessa, the Toronto deep learning startup that Square (now Block) bought in 2020; the model shares that company's name.^[1][9][11]

Announcement

Block published the Dessa page at block.xyz/dessa on 23 October 2026. The page's description calls Dessa "a research preview of a full-duplex speech model", and its headline line is "You learned to talk by listening. So did Dessa."^[1] Thurairatnam announced the model on X that day, writing: "We made the first AI model you can have a realistic conversation with that was trained entirely on speech! Dessa is a full-duplex speech-to-speech model that you can run on consumer GPUs. This is a GPT3 moment for speech."^[4] A second post linked the page and said the lab was "inviting AI safety researchers to explore Dessa".^[5] Dorsey's post followed less than two hours after Thurairatnam's first one, with a reply adding: "this is big (but also really really small...small enough to run on your macbook)."^[2][3]

Block frames the project as a reaction to cascaded voice assistants, which convert speech to text, process the text, and convert the answer back into speech. In the page's words, current voice AI "turns your voice into text and throws away an incredible amount of information text alone can't carry".^[1] The page argues that emotion, sarcasm, laughter, sighs and stress patterns change what a sentence means and are lost once speech is reduced to a transcript.^[1]

Technical description

The announcement page does not include a paper, model card, parameter count or architecture description, and CoinDesk reported that Block had not disclosed the model's size or benchmark results. The technical claims below come from the announcement page and are the company's own.^[1][8]

• Speech in, speech out. Block calls Dessa "a true speech-to-speech model" and says "everything it knows came from human speech."^[1]
• Full duplex. The model "listens and talks at the same time". Block says it learned from real conversation when to cut in and when to backchannel with short acknowledgements such as "yeah!", "got it!" and "ooh".^[1]
• Training data. Trained from scratch "on nothing but human conversation", with no text and no pretrained models. Block does not identify the conversation dataset or say how it was obtained.^[1][9]
• Data efficiency. Block says Dessa "learned language from hundreds of times fewer words" than large language models need, and that a spoken word takes "thousands of times more bytes" than the same word in text.^[1]
• Training compute. Block says it figured out how to train Dessa "on only one NVIDIA GB200 NVL72", a single rack-scale system.^[1]
• Inference. Inference "runs faster than real time" and the model "runs locally on a MacBook Pro M4 Max or better".^[1] Thurairatnam's announcement said it can run "on consumer GPUs".^[4]

The page also lists open research questions the team says a speech-only model raises: how efficiently such models learn language and facts compared with text models, how efficiently they learn new languages, and what kinds of in-context learning they can perform.^[1]

Demonstrations

The announcement page carries a set of two-track recordings (one track per speaker) that Block describes as live conversations with a person, with "nobody steers it": the model chooses what to say, when, and in what voice and personality. Subtitles were added afterward.^[1] Featured clips include:

For the GPT-Live-1 clip, Block notes that both voices are AI-generated and that "the developer of GPT-Live-1 is not affiliated with Block and has not reviewed or endorsed Dessa."^[1] A further 15 samples are listed under "More samples", and the page says more recordings would be shared on X.^[1] Thurairatnam separately posted a clip of a Block intern playing a guessing game with the model, saying Dessa "developed world knowledge and learned to reason purely from audio".^[6][7][9]

Availability and safety research access

Block positions safety as one of three reasons for the project, alongside curiosity and "better understanding between humans and AI". Its example is a person "in serious pain while a voice AI has no idea and just keeps going", which the lab calls "the safety problem" and says Dessa is "a step toward fixing".^[1] The page invites "a small group of AI safety researchers" to get early access "to find the risks we haven't anticipated"; the application form asks for a name, an institutional email address and an optional Google Scholar or blog link.^[1]

As of 30 October 2026 the page's Download and "Join the waitlist" buttons were placeholders: clicking them showed "Download coming soon." or "The waitlist opens soon.", and the page carried the note "Launch links are coming soon."^[1] The page did not state a license, a release date or whether weights would be published. CoinDesk reported on 24 October that Block had not disclosed the model's size, benchmark results or a public release date, and that its audio samples "remain company demonstrations rather than independent tests."^[8]

Team

The page describes the group as "a (very) small AI research lab at Block, founded by former members of Dessa, the Toronto deep learning startup". It says the team has built deep learning models since 2012 and generative speech models "for over eight years", and that it reports directly to Dorsey, whom it calls Block's "cofounder, Chairman and Block Head".^[1] The page is signed by Ragavan Thurairatnam as "Dessa lead" and credits, in alphabetical order:^[1]

Reception

The Verge's Jay Peters wrote that the chats Block shared "can occasionally be a little awkward and uncanny". He noted that in one clip shared by Thurairatnam the model initially mistakes the speaker for someone named "Ronaldo", and that a conversation about Christopher Nolan's films drifts into a tangent about Bryan Singer's Superman films. He wrote that in the exchange with GPT-Live-1, Dessa "comes across as more natural".^[7]

CoinDesk reported reactions split between amazement and concern about fraud. It quoted X users who praised how human the voice sounded, Binance founder Changpeng Zhao suggesting the model "might be very good for GiggleAcademy", his education project, and filmmaker Justine Bateman and journalist Shiv Aroor warning that humanlike voices could make scams and deception easier.^[8]

WinBuzzer noted that simultaneous listening and speaking is not new, pointing to Kyutai's Moshi and NVIDIA's PersonaPlex, both of which rely on language models trained on text. In its reading, where those systems combine spoken interaction with text-based foundations, Block is asking whether speech alone can supply both conversational behavior and knowledge. It also observed that the selected demos pair expressive speech with mistakes, and that Block does not identify its conversation dataset or how permission to use it was obtained.^[9]

Context: full-duplex and textless speech models

Block hedges its novelty claim: "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."^[1] Both halves of the idea have earlier research behind them.

dGSLM was presented as a generator of dialogue audio samples, and Moshi starts from a text language model. The closest earlier release is Standard Intelligence's hertz-dev (November 2024), an open full-duplex, audio-only base model that people could talk to live, published with weights trained both with and without text-model initialization; it was released as an untuned base model rather than a conversational system. Block's claim therefore rests on how realistic Dessa's live conversation is, and CoinDesk noted that Block's recordings are company demonstrations rather than independent tests.^{[8][13][14][17]}

Dessa (company)

Dessa was a Toronto machine learning company whose co-founders included Stephen Piron (co-CEO), Eric Kin-Ho Lee (co-CEO), Vince Wong (COO) and Ragavan Thurairatnam (head of machine learning).^[11] In May 2019 its engineers published audio clips from RealTalk, a speech synthesis system that cloned the voice of podcaster Joe Rogan; Dessa called it "the most realistic AI simulation of a voice we've heard to date" and said it would not release its research, model or datasets publicly until the public had first been made aware of the implications of speech synthesis.^[15] Square announced on 7 February 2020 that it had acquired Dessa to strengthen its machine learning and AI work, noting the team's work on deepfake detection; terms were not disclosed and the team stayed in Toronto.^[10][11] Square, Inc. announced on 1 December 2021 that it was changing its name to Block, Inc.; the change took effect on 10 December 2021.^[10][19]

Block's 2026 page says the new lab was founded by former members of that startup, and Thurairatnam's X profile describes him as a previous founder who "sold @dessa".^[1][16] CoinDesk likewise reported that the lab includes former staff of the Toronto startup Square acquired in 2020.^[8]

Continue reading

• GGPT-LiveGPT-Live is a family of full-duplex voice models developed by OpenAI for continuous spoken interaction.Mentioned 7 times above
• AAI safetyAI safety is the research and practice of preventing or reducing unacceptable harm from artificial intelligence systems.Mentioned 4 times above
• MoshiMoshi is a full-duplex speech-to-speech foundation model developed by Kyutai, a French nonprofit artificial intelligence research laboratory.Mentioned 3 times above
• VVoice AIVoice AI is the umbrella term for artificial intelligence systems that listen to, understand, and generate human speech.Mentioned twice above
• VVoice cloningVoice cloning is the use of machine learning to generate synthetic speech in the voice of a specific real person (the target speaker) from a sample of their recorded audio.Listed in See also
• TText-to-SpeechText-to-speech (TTS) refers to artificial intelligence systems that convert written text into natural-sounding spoken audio.Listed in See also

More from See also

• Speech Recognition

More to read

• Block, Inc.
• Standard Intelligence
• HuBERT
• Voice assistant
• Hume AI

Topics:AI ModelsConversational AISpeech & Audio AIVoice AI

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35 ^36 ^37 ^38 ^39Block, Inc. "Dessa - the Conversation Model trained only on speech". block.xyz, 23 October 2026 (accessed 30 October 2026). block.xyz/dessa
2. ^1 ^2 ^3Jack Dorsey (@jack). Post on X, 23 October 2026. x.com/...2113697902916751661
3. ^Jack Dorsey (@jack). Reply on X, 23 October 2026. x.com/...2113697906200883218
4. ^1 ^2 ^3Ragavan Thurairatnam (@RagavanThurair). Post on X, 23 October 2026. x.com/...2113671324400861645
5. ^Ragavan Thurairatnam (@RagavanThurair). Post on X, 23 October 2026. x.com/...2113691903866872251
6. ^Ragavan Thurairatnam (@RagavanThurair). Post on X, 23 October 2026. x.com/...2113704911749911028
7. ^1 ^2 ^3 ^4Jay Peters. "Jack Dorsey's Block made an AI model that talks to you like an awkward new friend". The Verge, 23 October 2026. theverge.com/...dorsey-block-dessa-ai-voice-speech
8. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Francisco Rodrigues. "Block's new voice AI sounds eerily human, sparking awe and scam fears". CoinDesk, 24 October 2026. coindesk.com/...-human-sparking-awe-and-scam-fears
9. ^1 ^2 ^3 ^4Markus Kasanmascheff. "Block's New Hyperrealistic Dessa Voice AI Learned From Human Speech Alone". WinBuzzer, 25 October 2026. winbuzzer.com/...-learned-from-speech-alone-xcxwbn
10. ^1 ^2Block, Inc. (then Square, Inc.). "Welcoming the Dessa Team to Square" (press release), 7 February 2020. block.xyz/...press-release-dessa-acquisition
11. ^1 ^2 ^3Meagan Simpson. "Toronto machine learning startup Dessa acquired by Square". BetaKit, 7 February 2020. betakit.com/...ng-startup-dessa-acquired-by-square
12. ^Kushal Lakhotia et al. "Generative Spoken Language Modeling from Raw Audio". arXiv:2102.01192, 2021. arxiv.org/...2102.01192
13. ^1 ^2Tu Anh Nguyen et al. "Generative Spoken Dialogue Language Modeling". arXiv:2203.16502, 2022. arxiv.org/...2203.16502
14. ^1 ^2Alexandre Défossez et al. "Moshi: a speech-text foundation model for real-time dialogue". arXiv:2410.00037, 2024. arxiv.org/...2410.00037
15. ^Katyanna Quach. "If you hear podcasting star Joe Rogan say something dumb, it may not be his fault - an AI has cloned his voice". The Register, 17 May 2019. theregister.com/...joe_rogan_ai
16. ^Ragavan Thurairatnam (@RagavanThurair). X profile (accessed 30 October 2026). x.com/RagavanThurair
17. ^1 ^2Standard Intelligence. "An Open Base Model for Conversational Speech" (hertz-dev). si.inc, 3 November 2024. si.inc/hertz-dev
18. ^OpenAI. "Introducing GPT-Live". openai.com, 8 July 2026. openai.com/...introducing-gpt-live
19. ^CNBC. "Jack Dorsey's Square changes corporate name to Block". 1 December 2021. cnbc.com/...square-changes-corporate-name-to-block-

Show all 19 references