Block says Dessa takes an unusual route to conversational voice AI: training a model from scratch on human speech, without text or pretrained models, rather than relying on a text-trained language model. That is a different claim from skipping a transcription stage at inference. Audio-native models that take speech in and put speech out already exist; Block's claim is about where the knowledge came from. Everything Dessa knows, Block says, arrived as sound.
The company frames it the way a child learns. Babies learn to talk without ever seeing a word written down, and Block wanted to see whether a model could too. The model shares its name with Dessa, the Toronto AI startup whose former researchers helped establish the small Block lab now developing it. Ragavan Thurairatnam is identified as "Dessa lead", and the lab reports directly to Jack Dorsey.
The training claim, and what is missing
Block says Dessa learned language from hundreds of times fewer words than large language models and was trained using one NVIDIA GB200 NVL72 rack-scale system, comprising 72 GPUs. It provides no audio-hour count, dataset size, comparison denominator, training duration or model parameter count. Those omissions prevent outsiders from evaluating its data-efficiency and training-cost claims.
The first-of-its-kind claim is Block's own, and Block itself hedges it with "as far as we know". None of that makes the claim false. It makes it unverified, which is a different thing: the speech-only training route has not been independently validated, and readers should hold the claim in that category until someone outside Block can test the model.
What the demos show, and what they do not
The demos are live recordings, and Block says nobody steers the model during them. Dessa picks what to say, when to say it and how, down to its own voice. The selected recordings illustrate the timing, interruptions and emotional expression Block wants Dessa to handle: laughing the moment something is funny rather than after the sentence ends, backchanneling while the other person talks, getting louder when asked to speak up, finishing a sentence when a speaker blanks on a film title, and answering a surprise announcement with what Block describes as teary gratitude. They do not establish that Dessa performs those tasks more reliably than other voice systems.
The same demos cut the other way. The Verge watched them and found a model that can be awkward and uncanny. In one conversation Dessa initially mistakes the speaker for somebody named "Ronaldo". In another, a discussion of Christopher Nolan films wanders into an unexpected tangent about Bryan Singer's Superman films. Those are ordinary failure modes for a conversational system, and they are useful precisely because Block left them visible. A demo reel is still a selection, though: a reel the maker chose to publish cannot show how often the model does any of this across an ordinary hour of conversation.
One practical claim is firmer than the superlatives, if Block's claim holds. Block says Dessa runs locally on a MacBook Pro M4 Max or better, faster than real time. That matters for voice in a way it does not for text: a conversation model that does not need a data center round trip can keep timing that network latency breaks, and it can in principle keep audio off someone else's servers.
Access is the tell
Block is not selling Dessa yet. There is no price, no API, no weights and no date. The first outsiders to get it are a small group of AI safety researchers, who apply with an institutional email. Everyone else joins a waitlist.
Block's stated safety rationale is about hearing what text throws away. A voice AI that cannot hear that someone is in serious pain will simply keep going, and Block presents Dessa - which natively hears emotion, sarcasm, a frustrated sigh, a voice cracking with sadness - as a step toward fixing that. A separate concern, and ours rather than Block's, is whether highly expressive speech generation could enable convincing impersonation or manipulation. The launch provides no evaluation of those risks. Giving safety researchers early access creates an opportunity to investigate them, but it does not establish that the risks have been addressed.
For everyone else, the honest position on 23 October is narrow. A well-resourced lab has produced recorded evidence that a speech-only model can hold a recognizable conversation, including the timing and emotion that text pipelines lose. It has not produced evidence that the model knows as much, reasons as well or fails as rarely as the text models it is implicitly compared with, because it has published nothing that would let anyone measure those things.
Sources
- [1] Block, "Dessa - the Conversation Model trained only on speech" (launch page, 23 October 2026) - primary source for the speech-only training claim and its "as far as we know" hedge, the "hundreds of times fewer words" and single GB200 NVL72 claims, full-duplex and faster-than-real-time inference, local running on a MacBook Pro M4 Max or better, the demo descriptions, Block's safety rationale, the lab's composition and the "Dessa lead" byline, and the safety-researcher early-access terms; read first-hand 23 October 2026Read source
- [2] The Verge, Jay Peters, "Jack Dorsey's Block made an AI model that talks to you like an awkward new friend" (23 October 2026, 19:32 UTC) - source for the "Ronaldo" misidentification and the Bryan Singer Superman tangent in the published demos, and for the awkward, uncanny characterization; read via a syndicated copy of the Verge articleRead source