I dug into Block's Dessa voice model after the "we're cooked" clip. It's a real announcement, but nothing is released or independently tested yet. Here's what's actually published.
After the Dessa clip went around here, a lot of the replies were "hoax until they release it" or "this is only good for scams." I teach ML and wanted to separate what Block has actually said from what people are assuming, so I went through the official page, the team's posts, the press coverage and the open full-duplex models it's being compared to. Everything below was checked on 27 Oct 2026, with sources at the bottom.
1. Is it a hoax?
The announcement is real. Block has a page at block.xyz [1]. The lab lead Ragavan Thurairatnam [2] and Jack Dorsey [3] posted about it on 23 Oct, and both The Verge [4] and TechCrunch [5] covered it.
What nobody outside Block has done yet is test it. I couldn't find a single independent hands-on report. TechCrunch says "there is no public access to Dessa at the moment" [5]. Every clip is one Block chose to publish. Block says "Every clip below is a live recording of a person talking to Dessa. Nobody steers it." [1] That's Block's description, though, not something anyone can check. The Verge, reacting to the demos, called some clips "a little awkward and uncanny" [4].
So "hoax" goes too far. "Unverified capability claims" is accurate.
2. What Block actually claims [1]
- "Audio in, audio out." It's a speech-to-speech model.
- It was trained "from scratch on nothing but human conversation... No text, no pretrained models."
- It's full-duplex: it "knows when to cut in and when to backchannel ('yeah!', 'got it!', 'ooh')."
- "Inference runs faster than real time." No latency number is given.
- "It runs locally on a MacBook Pro M4 Max or better."
- It was trained "on only one NVIDIA GB200 NVL72."
- It "learned language from hundreds of times fewer words" than LLMs. No number is given.
3. What is not published (as of 27 Oct)
There are no weights, API, paper, model card, parameter count, latency figure, benchmarks, list of languages or price [1][6]. I also searched GitHub and Hugging Face and found no Block or Dessa speech repo. The page has "Download" and "Join the waitlist" buttons, but in the page itself neither links anywhere, and a note says "Launch links are coming soon" [1]. The only outside access is an application form for "a small group of AI safety researchers" [1].
4. "Consumer GPUs" vs "MacBook Pro M4 Max"
The post title here came from Thurairatnam's post, which says it's "a full-duplex speech-to-speech model that you can run on consumer GPUs" [2]. Block's page names only "a MacBook Pro M4 Max or better" [1]. Both come from Block, and neither gives a RAM or VRAM figure. The "0.5 GB VRAM" number on some hardware-ranking sites is an automated estimate with no published model size behind it, and the same site also lists other memory figures [7]. I'd ignore it.
5. The Moshi to Sesame to Dessa lineage theory
I found nothing that supports it.
- The author lists for Moshi [8], Sesame CSM [9] and Dessa [1] don't overlap.
- Moshi starts "from a text language model backbone" [8]. Sesame CSM needs a Llama-3.2-1B backbone [10], and Sesame's own blog says CSM models only the text and speech content of a conversation, "not the structure of the conversation itself" [9]. Block claims "no pretrained models" [1]. If that's true, Dessa isn't a fine-tune of either.
- Block publishes no code or citations, so the theory can't be checked either way.
The lineage Block itself states: the lab was "founded by former members of Dessa, the Toronto deep learning startup" and has built "generative speech models for over eight years" [1]. Square acquired Dessa in February 2020 [11]. According to Dessa's 2019 post, the same company built RealTalk, a Joe Rogan voice clone [12]. I only saw [11] and [12] as search snippets, so treat those details as lightly checked.
6. Is "first speech-only model" new?
Block says "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]. The word "realistic" carries a lot there. Standard Intelligence's Hertz-dev (Nov 2024) was already described as an "8.5B parameter, full-duplex, audio-only base model," with a recipe that "learns linguistics both with and without text model initialization" [13].
7. Published latency for open full-duplex models (Dessa: not disclosed)
These figures measure different things, so don't rank them against each other.
|Model|Published latency|What it measures|Open weights|Source|
|:-|:-|:-|:-|:-|
|Moshi (Kyutai)|160 ms / 200 ms|theoretical / "in practice"|Yes|[8]|
|Hertz-dev (Standard Intelligence)|80 ms / 120 ms|theoretical average / real-world on one RTX 4090|Yes|[13]|
|PersonaPlex 7B (NVIDIA)|70 ms / 170 ms|Full-Duplex-Bench turn-taking: paper model / released checkpoint|Yes|[14]|
|Dessa (Block)|not disclosed|"faster than real time" only|No|[1]|
8. The scam question
Block frames its safety work around AI that can't hear emotion: "Imagine being in serious pain while a voice AI has no idea and just keeps going" [1]. Its page says nothing about voice cloning, impersonation, watermarking or telling callers they're talking to an AI. It says the model chooses "its own voice and personality" [1], but it doesn't say whether it can copy a voice from a sample. TechCrunch's guess is that the safety-researcher program "is likely to test out cases where a voice agent could trick people" [5]. That's the reporter's inference, not Block's.
The scam concern itself is well documented. In Dec 2024 the FBI warned that criminals "generate short audio clips containing a loved one's voice to impersonate a close relative in a crisis situation" [15]. In May 2025 it warned that "AI-generated voice messages" were being used to impersonate senior US officials [16]. My own read: those scams don't need full duplex. But natural interruptions and backchannels would remove the robotic pacing and lag that people are told to listen for. That's untested for Dessa.
9. What would actually settle it
- Downloadable weights or the promised local app, so people can measure memory use and latency on an M4 Max or a named consumer GPU.
- An unscripted public demo instead of chosen clips.
- A technical report covering parameter count, training data and evidence for "no text, no pretrained models."
- Full-Duplex-Bench numbers set beside Moshi and PersonaPlex.
- Published findings from the outside safety researchers.
Until then, I'm filing it as a real announcement with impressive but unverified demos.
For people who've run Moshi, Hertz-dev or PersonaPlex locally: what would you test first if Dessa weights dropped tomorrow, and which number would convince you the speech-only claim is real?
Sources (all accessed 27 Oct 2026)
1. Block, Dessa page: block.xyz
2. Ragavan Thurairatnam on X, 23 Oct 2026: x.com
3. Jack Dorsey on X, 23 Oct 2026: x.com
4. The Verge, 23 Oct 2026: theverge.com
5. TechCrunch, 24 Oct 2026: techcrunch.com
6. CellCog blog, 24 Oct 2026 (secondary): cellcog.ai
7. madebyagents, Dessa hardware page (automated estimate): madebyagents.com
8. Moshi paper, arXiv 2410.00037: arxiv.org
9. Sesame, "Crossing the uncanny valley of voice": sesame.com
10. Sesame CSM GitHub: SesameAILabs/csm
11. Square press release, 7 Feb 2020 (search snippet only): squareup.com
12. Dessa, "RealTalk," 15 May 2019 (search snippet only): medium.com
13. Standard Intelligence, Hertz-dev: si.inc
14. NVIDIA PersonaPlex paper, arXiv 2602.06053: arxiv.org
15. FBI IC3 PSA, 3 Dec 2024: ic3.gov
16. FBI IC3 PSA, 15 May 2025: ic3.gov
11 points · 3 comments
Mentions
I dug into Block's Dessa voice model after the "we're cooked" clip. It's a real announcement, but nothing is released or independently tested yet. Here's what's actually published. After the Dessa clip went around here, a lot of the replies were "hoax until they release it" or "this is only good for scams." I teach ML and wanted to separate what Block has actually said from what people are assuming, so I went through the official page, the team's posts, the press coverage and the open full-duplex models it's being compared to. Everything below was checked on 27 Oct 2026, with sources at the bottom. 1. Is it a hoax? The announcement is real. Block has a page at block.xyz [1]. The lab lead Ragavan Thurairatnam [2] and Jack Dorsey [3] posted about it on 23 Oct, and both The Verge [4] and TechCrunch [5] covered it. What nobody outside Block has done yet is test it. I couldn't find a single independent hands-on report. TechCrunch says "there is no public access to Dessa at the moment" [5]. Every clip is one Block chose to publish. Block says "Every clip below is a live recording of a person talking to Dessa. Nobody steers it." [1] That's Block's description, though, not something anyone can check. The Verge, reacting to the demos, called some clips "a little awkward and uncanny" [4]. So "hoax" goes too far. "Unverified capability claims" is accurate. 2. What Block actually claims [1] - "Audio in, audio out." It's a speech-to-speech model. - It was trained "from scratch on nothing but human conversation... No text, no pretrained models." - It's full-duplex: it "knows when to cut in and when to backchannel ('yeah!', 'got it!', 'ooh')." - "Inference runs faster than real time." No latency number is given. - "It runs locally on a MacBook Pro M4 Max or better." - It was trained "on only one NVIDIA GB200 NVL72." - It "learned language from hundreds of times fewer words" than LLMs. No number is given. 3. What is not published (as of 27 Oct) There are no weights, API, paper, model card, parameter count, latency figure, benchmarks, list of languages or price [1][6]. I also searched GitHub and Hugging Face and found no Block or Dessa speech repo. The page has "Download" and "Join the waitlist" buttons, but in the page itself neither links anywhere, and a note says "Launch links are coming soon" [1]. The only outside access is an application form for "a small group of AI safety researchers" [1]. 4. "Consumer GPUs" vs "MacBook Pro M4 Max" The post title here came from Thurairatnam's post, which says it's "a full-duplex speech-to-speech model that you can run on consumer GPUs" [2]. Block's page names only "a MacBook Pro M4 Max or better" [1]. Both come from Block, and neither gives a RAM or VRAM figure. The "0.5 GB VRAM" number on some hardware-ranking sites is an automated estimate with no published model size behind it, and the same site also lists other memory figures [7]. I'd ignore it. 5. The Moshi to Sesame to Dessa lineage theory I found nothing that supports it. - The author lists for Moshi [8], Sesame CSM [9] and Dessa [1] don't overlap. - Moshi starts "from a text language model backbone" [8]. Sesame CSM needs a Llama-3.2-1B backbone [10], and Sesame's own blog says CSM models only the text and speech content of a conversation, "not the structure of the conversation itself" [9]. Block claims "no pretrained models" [1]. If that's true, Dessa isn't a fine-tune of either. - Block publishes no code or citations, so the theory can't be checked either way. The lineage Block itself states: the lab was "founded by former members of Dessa, the Toronto deep learning startup" and has built "generative speech models for over eight years" [1]. Square acquired Dessa in February 2020 [11]. According to Dessa's 2019 post, the same company built RealTalk, a Joe Rogan voice clone [12]. I only saw [11] and [12] as search snippets, so treat those details as lightly checked. 6. Is "first speech-only model" new? Block says "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]. The word "realistic" carries a lot there. Standard Intelligence's Hertz-dev (Nov 2024) was already described as an "8.5B parameter, full-duplex, audio-only base model," with a recipe that "learns linguistics both with and without text model initialization" [13]. 7. Published latency for open full-duplex models (Dessa: not disclosed) These figures measure different things, so don't rank them against each other. |Model|Published latency|What it measures|Open weights|Source| |:-|:-|:-|:-|:-| |Moshi (Kyutai)|160 ms / 200 ms|theoretical / "in practice"|Yes|[8]| |Hertz-dev (Standard Intelligence)|80 ms / 120 ms|theoretical average / real-world on one RTX 4090|Yes|[13]| |PersonaPlex 7B (NVIDIA)|70 ms / 170 ms|Full-Duplex-Bench turn-taking: paper model / released checkpoint|Yes|[14]| |Dessa (Block)|not disclosed|"faster than real time" only|No|[1]| 8. The scam question Block frames its safety work around AI that can't hear emotion: "Imagine being in serious pain while a voice AI has no idea and just keeps going" [1]. Its page says nothing about voice cloning, impersonation, watermarking or telling callers they're talking to an AI. It says the model chooses "its own voice and personality" [1], but it doesn't say whether it can copy a voice from a sample. TechCrunch's guess is that the safety-researcher program "is likely to test out cases where a voice agent could trick people" [5]. That's the reporter's inference, not Block's. The scam concern itself is well documented. In Dec 2024 the FBI warned that criminals "generate short audio clips containing a loved one's voice to impersonate a close relative in a crisis situation" [15]. In May 2025 it warned that "AI-generated voice messages" were being used to impersonate senior US officials [16]. My own read: those scams don't need full duplex. But natural interruptions and backchannels would remove the robotic pacing and lag that people are told to listen for. That's untested for Dessa. 9. What would actually settle it - Downloadable weights or the promised local app, so people can measure memory use and latency on an M4 Max or a named consumer GPU. - An unscripted public demo instead of chosen clips. - A technical report covering parameter count, training data and evidence for "no text, no pretrained models." - Full-Duplex-Bench numbers set beside Moshi and PersonaPlex. - Published findings from the outside safety researchers. Until then, I'm filing it as a real announcement with impressive but unverified demos. For people who've run Moshi, Hertz-dev or PersonaPlex locally: what would you test first if Dessa weights dropped tomorrow, and which number would convince you the speech-only claim is real? Sources (all accessed 27 Oct 2026) 1. Block, Dessa page: block.xyz 2. Ragavan Thurairatnam on X, 23 Oct 2026: x.com 3. Jack Dorsey on X, 23 Oct 2026: x.com 4. The Verge, 23 Oct 2026: theverge.com 5. TechCrunch, 24 Oct 2026: techcrunch.com 6. CellCog blog, 24 Oct 2026 (secondary): cellcog.ai 7. madebyagents, Dessa hardware page (automated estimate): madebyagents.com 8. Moshi paper, arXiv 2410.00037: arxiv.org 9. Sesame, "Crossing the uncanny valley of voice": sesame.com 10. Sesame CSM GitHub: SesameAILabs/csm 11. Square press release, 7 Feb 2020 (search snippet only): squareup.com 12. Dessa, "RealTalk," 15 May 2019 (search snippet only): medium.com 13. Standard Intelligence, Hertz-dev: si.inc 14. NVIDIA PersonaPlex paper, arXiv 2602.06053: arxiv.org 15. FBI IC3 PSA, 3 Dec 2024: ic3.gov 16. FBI IC3 PSA, 15 May 2025: ic3.gov
Get a 3060 as a second card. Use CSM for an offline voice pipeline for your local models. Clone your own voices. It's insanely fast. SesameAILabs/csm
@pleometric reminds me a lot of Sesame: SesameAILabs/csm