Block announced Dessa on October 23, 2026, as a speech-only, full-duplex conversation research preview. Trained from scratch on human conversation audio, the model explores listening and speaking without a text-based training pipeline. Initial access is restricted, leaving its production integration and safety evaluation as open questions.
What Block means by speech-only and full-duplex
According to Block, Dessa was trained from scratch on human conversation audio, without text or pretrained models. It takes speech in and produces speech out. Full-duplex means it can listen while speaking, including overlapping acknowledgments and interruptions. These are Block’s technical claims, not independently reproduced findings.
The distinction matters: training on speech alone concerns how a model learned; full-duplex concerns how it interacts. Neither establishes factual reliability, business-system access or permission to execute transactions. For an appointment agent, hearing a correction while talking would be useful only if the corrected details reached the booking workflow accurately.
Announcement, researcher preview, API and weights are different milestones
As checked on November 1, 2026, Block’s project page offers recordings and invites a small group of AI safety researchers to request early access. It also displays download and waitlist wording, alongside a notice that launch links are forthcoming. That presentation should not be mistaken for a verified public download.
- Announcement: October 23 established the project publicly, not general availability.
- Research preview: Block is soliciting safety-research participation; an application is not guaranteed access.
- Public API: the checked official page supplies no public endpoint documentation or self-service API access.
- Downloadable weights: the checked page supplies no checkpoint package or accompanying distribution license.
Those access distinctions reflect the official page, not a claim that no private implementation exists.
CoinDesk’s October 24 reporting likewise described a research preview rather than a publicly available product, and reported that model size, benchmarks and a public release date had not been disclosed. Treat that as dated corroboration, not evidence of a subsequent launch.
The practical opportunity is conversational repair
Our assessment is that the most valuable business possibility is not a voice that sounds personable. It is an interface that could recover gracefully when a customer changes direction mid-sentence.
Consider a hypothetical caller saying, “Tuesday morning—actually, Wednesday afternoon.” A useful evaluation would check whether the agent stops its obsolete reply, captures Wednesday, confirms the intended time and prevents a Tuesday booking. This is a proposed test, not a reported Dessa capability or a test we performed.
The same distinction applies to expressive speech. A sympathetic tone should not substitute for escalation when a caller asks for a human. For business buyers, successful correction and safe handoff would be more consequential than a convincing laugh.
Block says inference runs faster than real time and that Dessa runs locally on a MacBook Pro with an M4 Max chip or better. Those claims do not supply measured end-to-end call latency, minimum memory, concurrent-call capacity or a downloadable deployment configuration. They are not grounds for purchasing hardware for Dessa today.
What would turn this preview into a deployment decision?
For a future business evaluation, request evidence across three separate layers: conversation behavior, task correctness and operational controls. Test interruptions and long calls; verify names, dates and corrections; then examine action permissions, audit records and human transfer. These are evaluation priorities, not announced Dessa features.
Local execution, if made available, would also require a concrete account of where recordings, logs and updates travel before a team could assess privacy. A fluent voice alone answers none of those questions.
The next meaningful milestone is therefore not another expressive recording. It is documented access with usable terms, reproducible evaluations and a clear deployment interface. Until then, Dessa is worth following as speech research—not treating as an available replacement for a business voice-agent stack.