ModelsArchitectures & capability
Block unveils Dessa, a full-duplex voice model trained only on recorded human conversation
Block's AI lab has released a research preview of Dessa, a speech-to-speech model trained from scratch on human conversation audio with no text, which it says handles interruptions, laughter and emotion and runs locally on a high-end MacBook.
From the The Verge · AI article of October 23, 2026 and cited sources: Dessa was trained on audio alone, with no text and no pretrained starting model. [2] [5] [6]
SOURCES[2] Block: Dessa block.xyz[5] arXiv: Generative Spoken Dialogue Language Modeling arxiv.org[6] arXiv: Moshi arxiv.org
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Dessa was trained on audio alone, with no text and no pretrained starting model. [2] [5] [6]
- Original source
- The Verge · AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- Studies
- arXiv 2203.16502 (Mar 2022) [5] date from arXiv ID
- When
- October 23, 2026 article published
- Covered
- Oct 24 edition
- Evidence
- 9 sources across 8 sites
Reach: not measured yet
Dessa was trained on audio alone, with no text and no pretrained starting model. It listens and speaks at the same time. Meta's dGSLM showed in 2022 that natural turn-taking and laughter can be learned from audio alone, but its output was not coherent enough to act as an assistant. Kyutai's Moshi made full-duplex conversation practical, but it relies on text tokens running alongside the audio. Dessa's claimed step is speech-only training that works well enough for live conversation. [2] [5] [6]
The efficiency figures come from Block alone. Block says Dessa needed hundreds of times fewer words than large language models, trained on a single NVIDIA GB200 NVL72 rack and runs on a MacBook Pro M4 Max or better. It gives no exact word count and no parameter count. [2] [8]
Evidence so far is demo clips and stated training figures.
Read the full assessment
If they hold up, speech-only training on one rack with laptop inference could make private on-device voice agents cheaper; factual reliability and instruction following without text training remain untested.
Executive brief
Block's small AI research lab says its new model, Dessa, can hold a realistic live voice conversation even though it learned only from recorded human speech. It saw no text and started from no pretrained model. Block also says Dessa needed "hundreds of times fewer words" than large language models, was trained on a single NVIDIA GB200 NVL72, and runs locally on a MacBook Pro M4 Max (Block). The Verge found the demo conversations mostly natural but sometimes awkward (The Verge). There is no paper, no benchmark and no independent test yet. For now, only Block's own demos support these claims.
What changed and event timeline
Square buys Dessa
Square, now Block, bought the Toronto machine-learning startup Dessa, which was known for open-sourced deepfake detection. Terms were not disclosed (;).
Meta shows text-free dialogue generation
Meta's dGSLM learned from 2,000 hours of two-channel conversation audio with no text. It produced laughter and overlapping turn-taking, but not a live conversation partner ().
Kyutai releases Moshi
Moshi listens and speaks at the same time ("full-duplex") with about 200 ms latency in practice. It relies on text through an "Inner Monologue" of text tokens, and its weights are public ().
First press reaction
The Verge called the demos sometimes "awkward and uncanny." In one, the model wrongly calls the speaker "Ronaldo"; in another, it goes off on a tangent about Bryan Singer's Superman films ().
Capabilities and access
- No version number or parameter count has been published (Tech Insider).
- What it does: takes audio in and gives audio out, listening and talking at the same time.
- Hardware: runs locally on a MacBook Pro M4 Max or better (Block).
Read the full section
- Model: "Dessa," research preview. No version number or parameter count has been published (Tech Insider).
- What it does: takes audio in and gives audio out, listening and talking at the same time. Block says it handles interruptions, short "yeah"/"got it" listening cues, laughter and emotion such as sarcasm or sighs, and runs faster than real time (Block).
- Hardware: runs locally on a MacBook Pro M4 Max or better (Block).
- Access: AI safety researchers can apply with an institutional email; everyone else can join a waitlist. There are no weights, no API and no pricing (Tech Insider; ArtRealm AI).
Technical analysis for researchers and developers
- No text anywhere in training. Moshi depends on text tokens running alongside the audio.
- Compute: one GB200 NVL72 for training. That is modest by frontier standards, if the claim is accurate (Block).
- Not yet published: the audio tokenizer, the architecture, the size and source of the training data, latency numbers, and any evaluation method.
Read the full section
- No text anywhere in training. Moshi depends on text tokens running alongside the audio. Dessa skips that step, so everything it knows had to come from audio alone (arXiv; Block).
- Compute: one GB200 NVL72 for training. That is modest by frontier standards, if the claim is accurate (Block).
- Not yet published: the audio tokenizer, the architecture, the size and source of the training data, latency numbers, and any evaluation method. Results cannot be reproduced.
- Block lists its own open research questions: how efficiently speech-only models learn language and facts, how quickly they pick up new languages, and whether they can learn from examples in the conversation (in-context learning) (Block).
Claims and evidence
- "World's first" realistic live conversational AI trained only on speech. Block itself qualifies this with "as far as we know" (Block).
- "Hundreds of times fewer words" than LLMs. No exact count is given (Tech Insider).
- Sounds more natural than OpenAI's GPT-Live-1. This rests on one conversation Block selected, and OpenAI is not involved (The Verge; Unite.AI).
Read the full section
- "World's first" realistic live conversational AI trained only on speech. Block itself qualifies this with "as far as we know" (Block).
- "Hundreds of times fewer words" than LLMs. No exact count is given (Tech Insider).
- Sounds more natural than OpenAI's GPT-Live-1. This rests on one conversation Block selected, and OpenAI is not involved (The Verge; Unite.AI).
- All of the above are Block's own claims. The demos were hand-picked, subtitles were added afterwards, and no outside group has tested the model (Unite.AI; ArtRealm AI).
Context and prior work
- Meta's dGSLM (2022) already showed that natural-sounding dialogue, laughter and turn-taking can be learned from audio alone.
- Kyutai's Moshi (2024) made full-duplex voice practical, but with text in the loop (arXiv).
- Dessa's claimed contribution is combining the two: speech-only training at a quality good enough for live conversation.
Read the full section
- Meta's dGSLM (2022) already showed that natural-sounding dialogue, laughter and turn-taking can be learned from audio alone. Its output was not coherent enough to act as an assistant (arXiv).
- Kyutai's Moshi (2024) made full-duplex voice practical, but with text in the loop (arXiv).
- Dessa's claimed contribution is combining the two: speech-only training at a quality good enough for live conversation.
- The team is made up of former Dessa startup staff led by Ragavan Thurairatnam, who reports to Jack Dorsey (Unite.AI).
Limitations, safety and contested findings
- Visible mistakes: the demos include getting a name wrong and wandering off topic (The Verge).
- Safety pitch: Block argues that today's voice AI is "nearly deaf" to emotion, for example missing that a user is in serious pain, and presents emotional awareness as a safety feature (The Verge).
- The counter-concern: commentators point out that a voice this emotionally expressive could also be used to manipulate people (Tech Insider).
Read the full section
- Visible mistakes: the demos include getting a name wrong and wandering off topic (The Verge).
- Safety pitch: Block argues that today's voice AI is "nearly deaf" to emotion, for example missing that a user is in serious pain, and presents emotional awareness as a safety feature (The Verge).
- The counter-concern: commentators point out that a voice this emotionally expressive could also be used to manipulate people (Tech Insider).
- Unconfirmed open-source claims: some social posts say Dessa is open-source. That has not been confirmed (Tech Insider).
Business and practitioner implications
- Nothing to build on yet.
- Cost and privacy, if the claims hold.
- Factual reliability is the main risk. That makes it unsuited to customer-service or payments work without testing.
Read the full section
- Nothing to build on yet. With no API, weights or benchmarks, voice teams should keep using their current setups: chained speech-recognition → LLM → text-to-speech pipelines, or Moshi-style full-duplex models.
- Cost and privacy, if the claims hold. Training on one rack and running on a laptop would make private voice agents that run on the device itself cheaper to build.
- Factual reliability is the main risk. A model without text training may be weaker on facts and following instructions, as Block's own unresolved points suggest. That makes it unsuited to customer-service or payments work without testing.
- Watch for outside evaluations from the safety researchers getting early access.
The source trail.
Sources (9)
Jack Dorsey’s Block made an AI model that talks to you like an awkward new friend
Article text retrieved; extracted text may omit tables or interactive elements.
theverge.com