Dessa learned to talk without reading. Now comes the harder test.

Evidence note: this analysis uses Block's announced claims, peer-reviewed research and independent reporting. ReadSigil has not tested Dessa, and Block has not released independent benchmarks or a detailed architecture report as of 28 October 2026.

Picture a customer calling a bank. The caller pauses, struggles to explain an unexpected transaction and tries to speak while the assistant is answering. A transcript can capture the words. It does not reliably capture the fear, the false start, the interruption or the moment a useful assistant should stop talking. Now picture the same fluid, attentive voice being used by a fraudster. That is the tension at the centre of Dessa.

On 23 October 2026, Block introduced Dessa, a research preview it describes as a realistic, full-duplex speech-to-speech AI trained from scratch on human conversation, with no text and no pretrained models. It says the model learned language using a single NVIDIA GB200 NVL72 system and can run locally on an M4 Max MacBook Pro or better. Those are company claims, not independently reproduced results. The Verge's reporting found conversations that sounded strikingly natural in places and awkward or tangential in others.

The most interesting question is not whether Dessa laughs convincingly. It is whether an intelligence trained from sound can preserve meaning, timing and social context without borrowing a text-trained brain—and whether those benefits survive real-world constraints.

01 / The technical change: from text in the middle to audio throughout

A conventional speech pipeline often sends microphone audio to automatic speech recognition (ASR), a language model, then text-to-speech (TTS). This is practical: text is easy to search, moderate, retrieve against policy, log and use in tools. But a transcription can erase hesitation, overlapping speech, vocal effort and the distinction between a reassuring sentence and a threatening one. Pipeline boundaries can also add latency.

Dessa is presented as a different arrangement: audio in, audio out, with continual listening and speaking. Full-duplex does not mean merely responding quickly. It means processing incoming speech while outgoing speech continues, including interruptions, acknowledgements and changes in conversational floor. Those are separate technical problems from language knowledge and factual correctness.

That last caveat matters. We do not know whether Dessa internally uses discrete acoustic units, continuous representations or another formulation; Block has not published enough to reconstruct its architecture. A direct audio interface is not a licence for analysts to invent the layers inside it.

02 / The breakthrough claim has predecessors

Speech-only is not a new research idea. The dGSLM research published in TACL in 2023 used 2,000 hours of conversational audio with no text labels and produced two-channel spoken dialogue including laughter and turn-taking. Its authors explicitly noted that coherent semantic dialogue remained a limitation. The Moshi project demonstrated real-time, full-duplex speech but began with a text-language-model backbone. NVIDIA's PersonaPlex added role and voice control on top of Moshi's architecture, which includes text-conditioned elements.

So Dessa should not be described as the first AI to generate conversation from sound, nor as the first duplex system. Block's narrower claim is a usable, realistic live conversation after speech-only training. That would be meaningful—provided independent tests establish the quality, generality and training boundary.

There is a subtle irony in the name. Square acquired the Toronto AI company Dessa in February 2020, bringing in a team associated with earlier generative voice research. The new model comes from a group with a longer history in the field than the launch-week excitement suggests.

03 / The numbers that look spectacular, and what they do not tell us

One NVIDIA GB200 NVL72 is an entire rack-scale system containing 72 Blackwell GPUs; it is not a commodity card. Training on one such system may still be notable relative to frontier language-model clusters, but one rack is not one GPU. The claim also excludes the details an engineer needs: model size, training duration, actual GPU utilisation, audio hours, dataset rights, codec, power consumption and total experiment budget.

Block says Dessa learned language from hundreds of times fewer words than large language models. That is not yet a comparable sample-efficiency benchmark. Spoken words, text tokens, audio frames, model objectives and prior knowledge are different denominators. Without the data recipe and a matched evaluation, a precise multiplier would give false confidence.

Likewise, “faster than real time” usually describes generation throughput. It does not by itself establish time to first audio, the delay after an interruption, p95 turn latency, memory footprint or performance in noisy conditions. Local inference on an M4 Max may suggest new privacy and economics options, but local deployment alone does not establish data retention, security or consent safeguards.

04 / Why “sounds human” and “understands humans” are different tests

An assistant can reproduce a pause or a compassionate tone without accurately understanding fear, deception, intent or risk. A September 2026 study of multiple full-duplex model families found that direct address and silence were much more reliable triggers for speaking than a false claim or a hazard. The models often waited for permission instead of intervening when content warranted it. That is a warning against treating natural turn-taking as reliable judgement.

For a financial-services assistant, the difference is operational. A voice that says “I understand” while missing a coercion cue is not safe simply because its prosody sounds kind. Equally, a system that sounds distressed is not proof that it experiences distress. Acoustic realism is a property of the interface, not evidence of sentience or correct reasoning.

05 / The future: five plausible moves if this holds up

1. Always-available local interfaces. Devices, vehicles and kiosks may converse without sending every utterance to a remote model. This depends on actual edge performance, battery use and access to tools—not a single laptop demo.
2. Voice-first accessibility. Systems that handle overlapping speech, changing pace and imperfect delivery could better serve users for whom typing is difficult. Testing must include disability and accent diversity rather than ideal demo voices.
3. Embodied and physical AI. Robots that must listen while acting cannot always wait for silence before responding. A duplex speech layer could eventually combine with vision, motion and safety control. Dessa itself has not demonstrated such integration.
4. Conversation as an enterprise data asset. Audio captures social signals text archives discard. That may make recorded interactions more valuable—and more sensitive. The hard problems will be consent, retention, provenance and unequal performance across languages.
5. A new impersonation economy. More natural machine conversation could lower the effort needed for convincing social engineering. That is a risk of the wider capability class, not evidence that Dessa currently offers voice cloning or fraud features.

A 2026 survey of voice-authentication threats documents how deepfake and spoofing attacks challenge both listeners and detection systems. That risk predates Dessa. A more fluid voice interface would make the boundary between a familiar-sounding caller and a verified identity even more important.

Text gave AI answers. Speech may give AI presence. Presence must never become permission.

06 / A test plan worth more than another demo

Before treating speech-only conversational AI as production-ready, I would ask for one evaluation protocol run on unreleased, consented conversations, with a text-mediated baseline, Moshi-class duplex baseline and human raters where appropriate. Report averages and failure tails, across accents, languages, network conditions and devices.

07 / If you were deploying it in a bank, here is the boundary

I would separate the conversational layer from the decision and action layers. The speech model could listen and speak. It could not independently prove who the customer is, approve a payment, change a limit or waive a control. Those actions need authenticated identity, scoped permissions, explicit transaction confirmation and an audit trail that is not merely another transcript of an agreeable voice.

• Prove the principal: authenticate the user through a separate factor and a secure session, not vocal resemblance or conversational familiarity.
• Bound the authority: tool actions must be policy-checked, least-privilege and revocable; payment instructions require transaction-specific consent.
• Protect the recording: minimise audio retention, specify consent and provenance, and test leakage across sessions.
• Expose the artificial: disclose that the caller or assistant is AI; do not let emotional realism obscure machine identity.
• Test the tail: include coercion, elder fraud, distress, code-switching, background speakers and failed authentication in red-team scenarios.

This is not a product endorsement of Dessa. The model is not generally available, and there is insufficient public evidence today to treat it as a deployable banking component.

The ReadSigil conclusion

Dessa's most interesting proposition is not that a machine can laugh. We already knew machines could generate laughter. It is that language competence and conversational timing may be learned from speech without making written text the centre of the system. If validated, that would widen the design space for AI—especially for local devices, people who interact through voice and environments where a second of delay matters.

The burden of proof now shifts from the demo to the measurement. Show the architecture. Release a testable model or controlled external access. Publish failure distributions, not just highlight clips. Document the speech data and its rights. And show what the system does when sounding sympathetic conflicts with doing the safe thing.

We should be excited by a system that can hold a conversation. We should be harder to impress when the conversation asks us to trust it.