A pause, an interruption or a quiet gasp can carry more meaning than a perfectly transcribed sentence. That is the promise behind Dessa, the voice-AI research preview from Block. The more useful question, though, is not simply whether a demonstration sounds human. It is what readers can actually try, and which claims still need independent testing.
Can you use Dessa today?
As of October 25, 2026, Block’s public Dessa page offers recorded conversations and invites a small group of AI safety researchers to seek early access. It is not presented as a general consumer app. Readers should distinguish listening to a sample from receiving access to the underlying model.
AsiaNews Iran reviewed the company’s research-preview page for this report. Its local-hardware claim does not amount to a public download: Block says the model can run on a MacBook Pro M4 Max or better, while access remains limited. A compatible computer alone therefore does not establish that its owner can install Dessa.
What does full-duplex speech mean?
Block describes Dessa as a speech-to-speech model that can listen and talk at the same time. The company says it trained the system from human conversations without text or pretrained models. Those are developer statements, not findings from an independent reproduction of the training process.
The practical distinction is turn-taking. A conversation is not always one complete message followed by another: people interrupt, acknowledge a point before it is finished, or change their meaning through emphasis. A system intended to handle those signals needs evaluation beyond whether its words form a fluent transcript.
- Timing: Does it stop or adapt when a person interrupts?
- Meaning: Does it retain the request rather than merely imitate its tone?
- Reliability: Does the behavior hold up outside a carefully chosen recording?
- Disclosure: Can a listener readily tell that one voice is generated?
Watch the official conversation demonstration
The recording below is Block’s “An anniversary surprise” demonstration. The track marked Human is the person speaking; the Dessa track is AI-generated speech. The waveform presentation and subtitles come from the developer. This is a published example, not an independent interview with the model or a live chat available on this page.
Block says its examples were recorded live and that subtitles were added afterward. The company also names Rayhane Mama, pictured in its project portrait, among the researchers on the Dessa team. Dessa lead Ragavan Thurairatnam signs the introduction, and the lab says it reports to Block cofounder Jack Dorsey.
The distinction between expression and experience matters. A generated gasp, laugh or sympathetic phrase is not evidence that a machine feels gratitude or understands distress as a person does. It may be a useful conversational behavior, but that is a different claim from consciousness, trustworthiness or suitability for sensitive advice.
What should the next evidence look like?
The invitation to safety researchers is an opening for outside scrutiny, not proof that the system has already passed every relevant test. Useful evaluation would describe the conditions, the failure cases and the limits, not just publish the most convincing exchanges.
- Check whether an access announcement is for researchers, developers or ordinary users.
- Separate a hardware capability claim from a released application, license and installation instructions.
- Look for independent tests covering interruptions, noisy surroundings, accents and misleading emotional responses.
Access restrictions and user control also shape other new AI products. Our guide to Claude Duet access and approvals examines that distinction in another setting; it is not a head-to-head comparison with Dessa.
For now, Dessa is best understood as a research preview with public recordings and limited early access. The demos make the design goal tangible; broader claims will need broader evidence. Which matters more to you in a voice assistant: natural timing, factual accuracy or unmistakable AI disclosure? Share your view in the comment box below the source, tags and related articles.