Block, the payments company run by Jack Dorsey, has unveiled Dessa, a voice model it says learned to hold a live conversation from human speech alone. The Dessa page on block.xyz describes a "true speech-to-speech model" trained "from scratch on nothing but human conversation. Sounds! No text, no pretrained models." It is a research preview with a waitlist, a set of recorded demos, and no paper, weights or benchmark tables yet. The Verge covered the announcement on October 23, 2026, calling some of the sample conversations "a little awkward and uncanny."
Here is what Block actually claims, how that differs from the voice assistants you use now, what the demos do and do not prove, and what to watch for next.
TL;DR: Dessa in one table
| Question | Answer |
|---|---|
| Who made it? | A small research lab inside Block that reports to Jack Dorsey, led by Ragavan Thurairatnam; it says it was founded by former members of the Toronto deep-learning startup Dessa |
| What is it? | A full-duplex, audio-in audio-out speech model, per Block |
| What is unusual? | Trained only on human conversation audio: no text, no pretrained model |
| Where does it run? | Block says locally on a MacBook Pro M4 Max or better, faster than real time |
| Training compute? | Block says one NVIDIA GB200 NVL72 |
| Can I use it? | Waitlist only; early access for "a small group of AI safety researchers" |
| What is missing? | A paper, weights, an API, latency figures, benchmark scores, a license |
What does "trained only on speech" actually mean?
Almost every voice assistant today is built on text. Speech is converted to text, a language model reasons over the text, and a synthesizer reads the answer aloud. Newer native audio models, such as the ones behind ChatGPT's voice mode (see our coverage of GPT-Live-1.1 and its new voices) and xAI's Grok Voice 3.0, skip some of those steps, but they still start from large language models pretrained on text.
Block's claim is stronger. The page says Dessa never saw written words: "everything it knows came from human speech." It argues this matters because converting speech to text discards information that text cannot carry, such as emotion, sarcasm, laughter, a frustrated sigh or a voice cracking. Its framing is a baby learning to talk without reading. If true, this is a different recipe from text-first models, and Block says that makes it "a different category with entirely new and interesting behaviors for researchers to discover."
Block adds that it believes Dessa is "the world's first AI you can have a realistic live conversation with that was trained only from human speech." The words "as far as we know" in the original soften this, and no independent party has checked it. Earlier research on speech-only language modeling exists, but not, to our knowledge, at a quality that lets a model hold open conversations, so the claim is plausible but unproven.
Full-duplex: why the demos sound different
Most voice assistants take turns. You speak, the system detects silence, then it answers. Dessa is described as full-duplex: it "listens and talks at the same time, like you do." Block says it learned when to cut in and when to backchannel with "yeah!", "got it!" or "ooh." In the published clips it laughs mid-sentence at a joke, gets louder when asked to speak up (which Block says nobody taught it), finishes a sentence when a speaker blanks on a movie title, and gasps "oh my god" before answering a surprise in what the page calls "teary gratitude."
Those behaviors are exactly where turn-based systems feel stilted, so this is the right place to push. Our coverage of Gradium's Phonon TTS and Cactus's Whistle speech-to-text shows how fast small on-device speech components are improving. Dessa goes the other way: a single model for the whole loop.
The efficiency claims
Block makes three quantitative-sounding claims, none with a published method:
- Less data. "LLMs need an insane amount of text. Dessa learned language from hundreds of times fewer words." Because the model never reads text, this presumably compares the number of words spoken in its audio training set to text-model corpora. Without a paper, we cannot see how the audio was counted.
- Small training hardware. It was trained on "only one NVIDIA GB200 NVL72." That is a single rack, large by hobbyist standards and modest next to the clusters used for frontier language models. Compare it with Meta's Prometheus supercluster plans to see the scale gap.
- Local inference. Faster than real time on a MacBook Pro M4 Max or better. This implies a model small enough to fit in laptop memory, but Block gives no parameter count.
The page also explains why speech is harder than text: a sentence spoken twice never produces identical data, and one spoken word takes thousands of times more bytes than the same word in text. Block's example is stress: "I didn't say she stole my CUDA code" changes meaning with each emphasized word. That is a fair statement of the problem, and it is why people expect speech-only learning to need far more data per unit of knowledge, which makes the "hundreds of times fewer words" line the thing most worth scrutinizing.
What the demos show, and what they do not
Block published about a dozen recordings plus "more samples", including conversations about Christopher Nolan's films, a surprise-trip story and a refund call. It states that nobody steers the model during these clips and that subtitles were added afterward. The Verge noted that in one clip the model mistakes the speaker for someone named "Ronaldo," and that it takes an unexpected tangent about Bryan Singer's Superman films.
One demo has Dessa talking with OpenAI's GPT-Live-1. Block says both voices are AI-generated, and adds a disclaimer that the developer of GPT-Live-1 is not affiliated with Block and has not reviewed or endorsed Dessa. The page describes Dessa as sounding more natural in that exchange. That is a subjective judgment on selected clips, not a head-to-head benchmark. For structured comparisons, look at the speech-to-speech indexes we cited for Grok Voice 3.0, and for transcription accuracy see the Microsoft MAI-Transcribe-2 report.
What the demos cannot tell you: how often Dessa fails, how it handles accents, noise or long sessions, whether facts it states are correct (the page admits it "knows a surprising amount, considering every fact it has come from raw sound", not that it is accurate), and what safety behavior it has. The Verge's description of "awkward new friend" suggests the experience is closer to human than polished.
Why Block says it built this
The team gives three reasons: "curiosity (aka fun), better understanding between humans and AI, and safety." The safety argument is specific. Block says AI is gaining power and independence while staying "nearly deaf to the emotion in our voices," and offers the example of a voice AI that carries on while a user is in serious pain. A model that natively hears distress could, in principle, react to it. The team is offering early access to a small group of AI safety researchers "to find the risks we haven't anticipated."
That framing fits a pattern at Block. Dorsey has recently positioned the company around open and agent-first tools, from the Buzz workspace for AI agents to his Open the Frontier argument for open weights. Whether Dessa will be open is not stated. Block has not announced a license, weights or a release date, so treat any assumption of openness as unconfirmed.
What to watch next
- A technical report. The key questions are the training data size and source, the model size, the architecture and how latency was measured.
- Independent testing. Safety researchers with early access, and anyone who gets off the waitlist, can test interruption handling, emotion recognition and failure modes.
- A release decision. Weights, an API, or a consumer product would change the competitive picture for the voice products from OpenAI, xAI and others.
- Safety findings. A model that expresses emotion convincingly raises questions about manipulation and attachment. The announcement does not describe guardrails.
- Ronaldo-style errors at scale. The misidentification in the demo is small, but it hints at the confabulation risk of any model that talks fluently without grounding.
What this means for builders
If you build voice products today, nothing changes this week: you cannot call Dessa. But it signals where research attention is going. First, native audio models that treat the voice itself as the data are gaining ground over text-mediated pipelines. Second, on-device, full-duplex conversation at laptop scale would lower costs and privacy barriers for voice apps, matching the trend we saw in small local speech components. Third, claims about "first" and "hundreds of times fewer words" deserve the same discipline as vendor benchmarks: ask for the paper, the eval set and the weights.
For now, the practical step is to join the waitlist if you do voice research, listen to the clips with the caveats above in mind, and wait for independent numbers.
Details are accurate as of October 23, 2026, based on Block's Dessa page and The Verge's report. Block may publish a paper or change terms later.
Related reading
- ChatGPT GPT-Live-1.1: new voices, six languages, usage limits by plan
- Grok Voice 3.0: speech-to-speech index and Starlink claims
- Gradium Phonon: 31M on-device TTS
- Cactus Whistle: 16MB on-device speech-to-text
- Microsoft MAI-Transcribe-2 streaming results
- Jack Dorsey's Buzz workspace
- Jack Dorsey's Open the Frontier