Block’s Dessa: A Speech-Only AI Preview, Not a Finished Voice Product

01 / The story

What is Block’s Dessa?

Dessa is a conversational speech model developed by a small AI research lab at Block. The official project page says it was trained from scratch on human conversation, without text or pretrained models. It describes audio input, audio output and full-duplex operation—listening and speaking at the same time. Those are Block’s disclosures, not an independently audited training recipe. [1]

TechCrunch’s October 24 report and the company posts it embeds place public promotion on October 23, 2026. We did not independently open those original social posts. The project page’s October 21 privacy-notice date should not be mistaken for a verified launch date. Our article is published and checked October 25. [1] [3]

The important change is the research approach, not proof of a finished voice assistant. A model that learns conversational timing from audio may preserve signals lost in transcription. That possibility still needs reproducible tests, release terms and a clearer technical report.

02 / The story

Key takeaways

• Block describes training from human speech alone, with no text or pretrained models.
• Full duplex means listening and speaking together, not waiting for neatly separated turns.
• The laptop claim is MacBook Pro M4 Max or better; detailed memory and installation requirements are not supplied.
• Faster-than-real-time inference is not a numerical zero-latency guarantee.
• Selected recordings do not establish public access, a license or independent superiority.

03 / The story

How is speech-only training different from a text-mediated voice assistant?

A conventional voice pipeline can transcribe spoken input, send the words to a language model and synthesize the resulting text as speech. That arrangement can work well, but the transcript may lose timing, emphasis, laughter and other nonverbal details. Block says Dessa learns directly from conversational audio rather than that text-mediated training path. [1]

“Trained only on speech” does not establish that the model has no tokenizer, audio codec or intermediate numerical representation. Sound must still be represented for computation. The official page does not supply a detailed architecture, parameter count, dataset inventory or reproducible training procedure, so exact internal modules remain unverified. [1]

The useful question is not whether text is always bad. It is which signals a particular system preserves and whether doing so improves a real task. A fluent voice can still provide the wrong facts; a less expressive system can still answer accurately. Conversational naturalness and factual reliability need separate evaluations.

04 / The story

What does full-duplex conversation mean?

Full duplex means both sides can listen and speak concurrently. For an AI, that could allow backchannels such as “got it,” a response to an interruption, or a change in tone before the other speaker finishes. Block describes these behaviors in its selected recordings. [1]

That is different from a system that waits for an end-of-speech detector before producing any response. But talking over someone is not automatically good turn-taking. The appropriate moment to interrupt depends on the conversation, noise, language and user needs. Those conditions are not measured in the public page.

The recordings can be found on the [official Dessa project page](https://block.xyz/dessa). They remain with the original publisher; Qlauson has not downloaded or rehosted them. We reviewed the descriptions rather than listening to or independently timing the clips, so this is not an audio-quality review. [1]

05 / The story

What do the official demonstrations show?

Block says the clips are live conversations without someone steering the model, with subtitles added afterward. Descriptions emphasize laughter mid-sentence, helping someone recall a movie, emotional delivery and volume changes. These are the publisher’s selected examples, not a randomized sample of ordinary user sessions. [1]

The page also includes conversations with another AI and explicitly states that the other model’s developer is unaffiliated and has not reviewed or endorsed Dessa. A side-by-side recording is not necessarily a controlled comparison: prompts, context, microphone conditions and selection all influence what the audience hears. [1]

We do not turn emotional prosody into a claim that the model feels empathy or understands a person’s state reliably. Nor do these samples establish robust memory, multilingual coverage, accessibility, noise tolerance or long-conversation performance. Those are important open questions for subsequent research.

06 / The story

Is Dessa really zero-latency AI?

No numerical zero-latency result is established here. Block’s actual statement is that inference runs faster than real time. That means computation can keep up with the audio stream under its described setup; it does not eliminate capture, buffering, scheduling or response generation delays. [1]

A useful latency report would state the hardware, audio conditions, first-response delay, interruption-response time and how delays vary across many sessions. The official page does not provide those measurements, so we do not claim a particular number of milliseconds. [1]

For a real application, the perceived result also depends on whether the agent responds at the right moment and understands the interruption. A fast but inappropriate interruption can feel worse than a slightly slower clear answer. That is a design consideration, not a measured Dessa result.

07 / The story

Can Dessa run locally on a laptop?

Block claims local inference on a MacBook Pro M4 Max or better. The statement does not identify required RAM, exact GPU configuration, precision, memory footprint, power use or installation steps. It should not be generalized to every MacBook, phone or consumer graphics card. [1]

Training and inference are different workloads. Block says it trained Dessa on one NVIDIA GB200 NVL72. NVIDIA describes that as a rack-scale system with 72 Blackwell GPUs and 36 Grace CPUs—not one GPU. A claim about training on one rack is not a claim that the model was trained on an ordinary laptop. [1] [4]

Local inference could be attractive for responsiveness or data control, but feasibility is not a privacy policy. Until actual software and requirements are available, this remains a hardware claim rather than an installation tutorial or a verified offline workflow.

08 / The story

What can users download, and what does it cost?

The reviewed official page shows Download and Join the waitlist labels, while its extracted page content also says launch links are coming soon. It invites a small group of AI safety researchers to request early access. TechCrunch reported no public access on October 24. Those signals support research-preview framing, not a promise that everyone can use it immediately. [1] [3]

No public weights, inference API documentation, model license or price was established on the official page. That is a scoped research finding, not proof that no private arrangement exists. We do not call Dessa free, open source, open weight or commercially available without the corresponding release material. [1]

09 / The story

Who built Dessa, and how does it relate to the older company?

The project page names Ragavan Thurairatnam as the Dessa lead and describes a lab founded by former members of the Toronto startup Dessa, reporting directly to Jack Dorsey. Square announced its acquisition of that earlier company on February 7, 2020. The historical acquisition and the new model preview are separate events. [1] [2]

Earlier work by the company on deep learning or deepfake detection does not establish the new model’s architecture, open-source status or license. Likewise, Block’s financial products do not automatically share this project’s capabilities or data terms. Avoid filling current documentation gaps with facts from an older business. [1] [2]

010 / The story

What privacy and safety questions remain?

The official page includes a privacy notice effective October 21, 2026 covering website interactions. It describes collection of contact, device and usage information, cookies and sharing with service providers. It does not clearly establish an inference-specific promise that spoken conversations are never uploaded, logged or used for training. Local execution alone cannot answer those questions. [1]

For a future deployment, ask about audio retention, consent, training-data rights, telemetry, deletion and whether the system works without network access. Keep personal or business-sensitive conversations out until the relevant terms are clear. These are Qlauson’s review recommendations, not findings from an installed model.

Humanlike speech may encourage emotional overtrust or make synthetic speech harder to recognize. However, we did not establish that Dessa can clone an arbitrary person’s voice. We therefore do not present celebrity imitation, speaker enrollment or cloning controls as confirmed features. The early-access safety invitation is a request for investigation, not proof that the risks are solved. [1]

011 / The story

Why is the preview worth following?

Dessa offers a distinctive research direction: learning language and conversational behavior together from sound rather than treating voice merely as a wrapper around text. That is meaningful even if a finished product, broad factual competence and independent safety evaluation remain unresolved.

For developers and businesses, the next useful evidence is a technical report, reproducible latency and reliability tests, training-data disclosures and an actual release package with terms. For ordinary users, the question is simpler: does it respond appropriately and accurately when the conversation stops looking like a carefully selected demo?

This is a real Block research preview, not a proven replacement for existing assistants. We can take its approach seriously without declaring it the most natural, safest or best voice model. Those comparisons require evidence the current announcement does not supply.