Block's Dessa Voice Model Learns From Speech Alone

Introduction

On October 23, 2026, Block published a page for Dessa, a speech-to-speech conversation model built by a small AI research lab inside the company. Block's description is direct: "We trained Dessa from scratch on nothing but human conversation. Sounds! No text, no pretrained models." The model takes audio in and produces audio out, and it is full-duplex, meaning it listens and talks at the same time.

Two framings in Block's own copy are worth marking as claims, not findings. The page calls Dessa "the most authentic full-duplex speech model ever built," and 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." Dessa lead Ragavan Thurairatnam posted on X the same day that "this is a GPT3 moment for speech." These are the team's characterizations. The page offers recordings as evidence and, as of October 25, does not offer benchmarks, a paper, weights or public access.

Feature Overview

1. Speech in, speech out, with no text stage. Block argues that conventional voice AI "turns your voice into text and throws away an incredible amount of information text alone can't carry." Dessa is described as a true speech-to-speech model whose knowledge "came from human speech." Block says it learned a real-time conversation "simply by listening."

2. Full-duplex turn-taking. Because it listens while speaking, the model is said to have learned when to cut in and when to backchannel with sounds such as "yeah!", "got it!" and "ooh." In one demo clip it laughs mid-sentence at something funny, and Block says it gets louder when asked to speak up: "Nobody taught it that."

3. Emotion in the audio itself. Block says the model natively hears and speaks "emotion, sarcasm, laughter, a frustrated sigh, a voice cracking with sadness." TechCrunch notes that other speech models, such as ElevenLabs v4 and Fish Audio Drama 3, use voice tags to drive emotion, and writes that "it's not clear if that emotion is being steered by some kind of tags" in Dessa.

4. Local, faster-than-real-time inference. Block states that inference runs faster than real time and that Dessa "runs locally on a MacBook Pro M4 Max or better."

5. A small training footprint. Block says it trained Dessa on "only one NVIDIA GB200 NVL72," which TechCrunch paraphrases as a single NVIDIA GPU rack. The page also says Dessa "learned language from hundreds of times fewer words" than LLMs need, without giving an exact figure, and argues that speech is harder than text because a spoken word takes "thousands of times more bytes" than the same word typed.

What the page does not say.

Usability Analysis

Today there is little to use. TechCrunch reports there is no public access to Dessa. The page's waitlist and download buttons are marked as coming soon ("The waitlist opens soon"), and it carries a request form for a small group of AI safety researchers who will get early access "to find the risks we haven't anticipated." Block frames safety as one of three reasons for the project, alongside curiosity and better understanding between humans and AI. Its safety argument is that voice AI is "nearly deaf to the emotion in our voices," and it gives the example of a voice assistant that cannot tell the user is in serious pain.

What can be assessed is the demo material. Block says every clip is "a live recording of a person talking to Dessa," that "nobody steers it," and that subtitles were added afterward. One recording has Dessa talking with OpenAI's GPT-Live-1. A disclaimer on the page says GPT-Live-1's developer is not affiliated with Block and has not reviewed or endorsed Dessa.

Press coverage of the same clips is mixed. The Verge describes the conversations as occasionally awkward and uncanny. In one, the model initially mistakes the speaker for somebody named "Ronaldo." In another, a discussion of Christopher Nolan's films, Dessa goes on a tangent about Bryan Singer's Superman films. TechCrunch notes that Square, another Block business, already uses voice AI to handle ordering.

Because the clips were chosen and published by Block, they show what the model can do at its best, but they do not show how often it succeeds or fails.

Pros and Cons

Pros

• Full-duplex design with backchannels and mid-sentence laughter.
• Local runtime on a high-end MacBook Pro, with faster-than-real-time inference per Block.
• A stated safety rationale and an early-access route for outside safety researchers.
• A small team that says it has built generative speech models for over eight years.

Cons

• The evidence is Block-selected demo recordings, with no benchmark or paper on the page.
• No public access, weights, license or release date.
• Parameter count and speech-data volume are undisclosed, so the "hundreds of times fewer words" comparison cannot be checked.
• Published coverage notes lapses such as the "Ronaldo" misidentification and an off-topic tangent.

Outlook

Block poses open questions on the page itself: how efficiently speech-only models learn language and facts compared with text, how efficiently they learn new languages, and what sorts of in-context learning they can do. Those questions can only be answered with outside access, which is why the safety-researcher program matters more than any single clip.

For Block, TechCrunch observes that a voice model that runs on-device could support more business and consumer experiences at lower cost, which is the outlet's reading and not a Block statement. The team describes itself as "a (very) small AI research lab at Block, founded by former members of Dessa, the Toronto deep learning startup," reporting directly to Jack Dorsey. Whether Dessa becomes a product, a research release or neither is not stated.

Conclusion

Dessa is an interesting research claim with a thin public evidence base. The idea of training a conversational model on speech alone, with no text and no pretrained models, is clearly stated, and the demos suggest natural timing and emotional range. But the superlatives are Block's own, and nothing on the page lets an outsider verify efficiency, quality or safety. We rate it 3 out of 5 for now, with the rating dependent on evaluations, a paper or access arriving. It is most relevant to speech researchers, voice-product teams and AI safety researchers.

Editor's Verdict

Block's Dessa Voice Model Learns From Speech Alone is a workable proposition that fills a clear gap, even if it doesn't fundamentally change the landscape.

The strongest case for paying attention: a full-duplex design lets the model listen and speak at once, including backchannels such as yeah and mid-sentence laughter. That alone raises the bar for what readers should expect in this space. Reinforcing that, a local runtime on a MacBook Pro M4 Max or better, with inference faster than real time, keeps the model on the user's own machine according to Block — practical value rather than just headline appeal. The broader signal worth registering is straightforward: the no-text, no-pretrained-models claim is the central differentiator, and it rests on Block's own description because no paper or evaluation backs it yet. On the other side of the ledger, one constraint is real rather than a marketing footnote: the evidence is a set of demo recordings that Block selected and subtitled, with no benchmark, evaluation or paper on the page. It should factor into any serious decision. Layered on top of that, a model with no public access, weights, license or release date cannot be tested or reproduced by outside developers — which narrows the set of teams for whom this is an obvious yes.

For ML researchers, technical leads, and readers tracking the underlying science behind new capabilities, the smart move is to track its trajectory and revisit once the rough edges are filed down. For everyone else, the safer posture is to monitor coverage and revisit once the use cases that matter to your team are demonstrated in the wild.