Block unveils Dessa, a speech-only AI model that listens while it talks

AI models · October 23, 2026

Block has introduced Dessa, a conversational AI model that the company says learned entirely from human speech, without text or a pretrained model. It can listen and speak simultaneously, a design intended to make interruptions, laughter and changes in tone part of the conversation rather than noise to discard. Block’s announcement

The debut is a research preview, not a general release. Block is offering early access to a small group of AI safety researchers. Its announcement does not provide public model downloads, API pricing or a date for broader availability.

The experiment is in how it learns

Voice AI often has to manage two distinct jobs: understanding what was said and deciding how to respond. Systems that pass speech through a text transcript can lose details such as overlapping voices and emotional inflection. Kyutai described these limitations in its 2024 Moshi research paper, which introduced a full-duplex speech-text model built from a text-language backbone.

That history matters when assessing Dessa’s novelty. Simultaneous listening and talking is an established research direction. Block’s more distinctive claim is that Dessa acquired its conversational abilities from sound alone, without that text foundation. Whether this training approach delivers better results across languages, accents and difficult conversations remains an open question.

What the demonstrations establish

Block has posted recordings it describes as live, unsteered conversations, with subtitles added afterward. The examples highlight timing and expression: joining a speaker mid-thought, laughing during a conversation and changing vocal volume. These are company-selected demonstrations, not a controlled comparison. Listen to Block’s samples

The Verge’s October 23 report notes both a largely natural conversational rhythm and awkward moments, including a sample in which Dessa addresses its conversation partner by an unexpected name. That is a useful reminder that convincing delivery and reliable answers are separate qualities.

For an eventual voice assistant, handling a user’s interruption could be as important as producing a fluent sentence. But a polished clip cannot show how consistently a system does that in noisy rooms, with unfamiliar speakers or over a long session. Comparable tests would be needed before treating Block’s claims of exceptional realism as an established lead.

Local operation is a claim, not an available download

Block says Dessa runs locally on a MacBook Pro with an M4 Max chip or better and that training used one NVIDIA GB200 NVL72 system. Those statements do not establish an end-user memory requirement, a total training cost or a reproducible deployment recipe; the launch page does not supply those details. Technical claims

Developers therefore have no public release to benchmark or integrate from this announcement. A future model card, downloadable package and clear licensing terms would make the local-computing claim more actionable.

Safety testing comes before wider access

Block frames better recognition of vocal emotion as a safety motivation and asks outside researchers to investigate risks it may have missed. Researcher-access announcement

That rationale should not be mistaken for evidence of safe deployment. A system can sound sympathetic while misunderstanding a request. Useful independent testing would examine factual reliability, responses to distress, inappropriate persuasion and whether users can reliably interrupt or stop it.

Dessa’s significance today is the new speech-only training claim and the behavior shown in its demonstrations. The next milestone is evidence beyond those samples: outside evaluation and a clear account of who can use the model, on what hardware and under which safeguards.

Sources and reporting note

Reporting is based on Block’s original announcement, The Verge’s October 23 coverage and Kyutai’s Moshi paper. PromptsZone has not independently tested Dessa. Training, hardware and capability statements are attributed to Block; broader implications are analysis.

Illustrative photo: Wilfredor, “Yeti Microfone,” via Wikimedia Commons, September 23, 2022. Released under the CC0 1.0 Universal public-domain dedication. Display resized. The photograph does not show Block’s equipment or imply photographer endorsement.