Block has unveiled Dessa, a research-stage artificial intelligence model designed to listen and speak at the same time, without using text as an intermediate step. Announced on October 23 and examined in new reporting over the weekend, the system was trained from scratch on recorded human conversations, according to the company. Its goal is to capture the timing, emphasis, hesitation, laughter and other signals that conventional speech-to-text pipelines can lose.
The Dessa voice AI announcement matters because speaking naturally with software is a different problem from producing fluent written answers. Block says its model can respond to interruptions, vary its delivery and run locally on a suitably powerful laptop. But this is still a restricted research preview: there is no general public download, published model size, independently verified benchmark or confirmed commercial release date. Those distinctions are essential for anyone evaluating the technology for a product or business. TechCrunch
What Is Dessa, and What Did Block Announce?
Dessa is a speech-to-speech model developed by a research group inside Block, the financial technology company behind Square and Cash App. The team announced the system on October 23, 2026, and published recordings intended to demonstrate how it handles natural conversation. Independent coverage on October 24 and 25 brought attention to both its apparent fluency and the risks of increasingly convincing synthetic speech.
Block says the model learned from human conversation audio alone, without text training data or a pretrained language model. That is an unusually ambitious design choice. Many production voice systems have historically relied on a pipeline that transcribes a user’s words, passes the transcript to a language model and synthesizes the answer as speech. More recent systems also use direct audio modeling, so speech-to-speech itself is not a new invention. The distinctive claim here is the combination of speech-only training, real-time two-way interaction and local inference.
The company describes Dessa as full-duplex: it can process incoming speech while producing outgoing speech. In practical terms, a person can interrupt it, laugh during its response or make a brief acknowledgment without necessarily forcing a rigid stop-and-start exchange. The company’s demonstrations include guessing games and conversations with expressive shifts in delivery. They are selected demonstrations, not a controlled study of reliability. The Verge
Block’s official Dessa research announcement provides the company’s description of the model and its intended capabilities.
How Dessa Voice AI Differs From Traditional Voice Assistants
The easiest way to understand the engineering challenge is to compare a phone conversation with a typed chat. In text, the words arrive in discrete messages. In speech, the meaning also depends on pauses, stress, pitch, volume, overlapping voices and the moment a person decides to respond.
Consider the sentence, “I thought you finished the report.” Emphasizing I, you or finished can convey different expectations. A transcription system may preserve every word while losing the emphasis that explains the speaker’s intent. A laugh can soften criticism; a long silence can signal uncertainty; a quick interruption can mean enthusiasm rather than impatience.
A conventional cascade typically has three stages: automatic speech recognition, language processing and text-to-speech generation. Each stage can be optimized, and some modern products perform remarkably well. However, separating them may introduce latency and discard acoustic information unless the architecture explicitly preserves it. A direct audio model can, in principle, learn relationships between speech content and how it is delivered.
That does not automatically make its answers more accurate. Recognizing sarcasm is not the same as verifying a fact. Sounding sympathetic is not the same as understanding a person’s emotional state. A convincing conversational rhythm can make mistakes harder to notice, which is why evaluation must include factual reliability and appropriate behavior alongside naturalness.
What “full-duplex” really means
Full-duplex describes simultaneous communication, not human-level intelligence. Ordinary telephone calls are full-duplex because both participants can talk and hear each other at once. An AI voice system needs additional capabilities: detecting when the user is interrupting, deciding whether to yield, recognizing background sounds and managing overlapping audio without entering a feedback loop.
In a practical customer-service scenario, a caller might say, “I need to change my delivery address—actually, wait, I also have a question about the order.” A responsive agent should notice the correction, stop pursuing the first instruction and ask what the caller wants to handle first. That is a more demanding interaction than waiting for a complete sentence, converting it to text and reading back a prepared answer.
Why Speech-Only Training Is the Important Experiment
Block’s researchers say they trained Dessa from scratch on human speech, with no text tokens and no pretrained model. The company also claims the system acquired conversational knowledge from far fewer words than text-based large language models typically use. Those comparisons remain vendor assertions because the underlying dataset size, model architecture, parameter count and methodology have not been fully published.
The experiment raises a valuable scientific question: how much of conversational competence can a model learn from the structure of spoken interaction itself? Human beings learn language in environments rich with sound, gestures and social feedback, rather than by reading a large text corpus first. Audio-only learning is not equivalent to human learning, but it tests whether useful language representations can emerge from a different training signal.
The Dessa voice AI approach could be particularly relevant to applications where the sound of speech is part of the task. Examples include interactive language practice, accessibility tools, games and customer conversations that require careful turn-taking. Yet the absence of a text-centered architecture may create its own integration challenges. Developers still need ways to retrieve accurate business records, invoke tools, maintain audit logs and evaluate what a model said. Block has not established that Dessa currently solves all of those requirements.
Training compute is not the same as inference hardware
Block says it trained Dessa on a single NVIDIA GB200 NVL72 system. That is a rack-scale AI computing platform, not a single consumer graphics card. The company separately says the trained model can run locally on a MacBook Pro equipped with an M4 Max processor or better, with inference faster than real time under its reported conditions.
These are two different claims: one concerns the computing infrastructure used to build the model; the other concerns hardware needed to execute the finished model. Neither reveals the model’s parameter count, memory footprint, average latency under noisy conditions or energy consumption. Independent tests will be necessary before developers can compare its efficiency with other voice architectures. CoinDesk
Can Dessa Voice AI Run Locally on a Laptop?
According to Block, yes: the model can run on an appropriately configured MacBook Pro with an M4 Max chip or better. Local execution could be meaningful for products that need quick responses, work with unreliable connectivity or seek to limit the amount of live audio sent to remote servers.
However, “runs locally” does not mean “available to install.” As of October 26, Block has not published general-purpose model weights, an API, consumer application, pricing or a public release schedule. The company has invited a small group of AI safety researchers to request early access, while broader launch information remains pending. People should use the official product page rather than trust third-party sites advertising downloads or invitation codes. Voice AI Space
Even after a public release, a local deployment would need to be evaluated on more than whether it launches. Real-time audio systems can become unstable when microphones pick up speaker output, when several people talk at once or when the environment is noisy. Memory use, battery drain, heat, language support and latency on sustained sessions would all matter to product teams.
Does local processing guarantee privacy?
No. A locally executed model may reduce the need to transmit audio to a cloud provider, but the surrounding application could still upload recordings, diagnostics, transcripts or account data. Whether a product protects user information depends on the entire software design, permissions and retention policy—not solely the model’s ability to run offline.
That distinction is especially important for workplace calls and accessibility tools, where recordings may contain sensitive information. Organizations should ask whether audio is stored, whether third parties receive it, whether recordings are used for model improvement and how users can delete data. Block has not yet provided the documentation needed to assess a hypothetical commercial Dessa integration.
Dessa Voice AI and the Risks of Humanlike Speech
More natural conversation can improve accessibility and usability. It can also make impersonation and manipulation more convincing. A system that pauses, laughs and responds to interruptions may be harder for a listener to distinguish from a person than a stiff prerecorded message.
Coverage from TechCrunch, The Verge and CoinDesk highlights both the natural-sounding demonstrations and concerns about fraud. Those concerns are reasonable in the broader synthetic-voice market, but they should not be mistaken for evidence that Dessa itself has been used in scams. There is no such confirmed claim in the available reporting. TechCrunch
The Dessa voice AI preview is being opened first to a small group of safety researchers. That approach may help identify unsafe responses, misleading behavior or ways the model could be misused before broader distribution. It does not, by itself, prove that the technology is safe. Meaningful assurance would require published testing criteria, documented mitigations and repeatable outside evaluation.
What responsible voice products should do
Businesses adopting humanlike voice agents should clearly disclose when a caller is interacting with AI, avoid presenting synthetic speech as a particular real person without authorization, and provide a reliable way to reach a human representative. Financial instructions and account changes should require independent verification, regardless of how familiar or credible the voice sounds.
Consumers should likewise avoid treating vocal realism as proof of identity. If an unexpected caller asks for money, credentials or urgent action, independently contact the supposed sender through a trusted channel. These are sensible precautions for synthetic voice technology generally, not a response to a verified Dessa-specific incident.
Where This Technology Could Be Useful
One plausible application is customer support. An assistant that manages interruptions well could make order inquiries and appointment scheduling less frustrating. But it would still need reliable connections to databases, a clear escalation process and safeguards against inventing policies or prices. Natural voice alone is not a complete service system.
Education is another possibility. A language-practice tool could offer learners a conversation with more realistic timing and expressive feedback. It would need to distinguish between friendly dialogue and accurate instruction, however, and developers would need to evaluate language coverage, accents and accessibility needs.
Games and interactive entertainment could benefit from characters that respond fluidly rather than waiting for a scripted turn. Assistive technology might also gain from systems that recognize when a user is struggling to complete a sentence. In sensitive contexts, those capabilities should be designed to support users without claiming to diagnose their emotions or health.
For Block, potential relevance to its broader commerce ecosystem is obvious. The company already operates tools used by merchants, and voice interfaces could eventually support business interactions. But Block has not announced that Dessa is being integrated into Square, Cash App or any particular commercial product. Treating such integrations as confirmed would be premature.
What Independent Testing Still Needs to Establish
The most important unanswered questions are measurable. Can the model maintain coherent conversation over long sessions? How often does it interrupt at the wrong time? Does it reliably understand speakers with different accents, speech impairments or noisy backgrounds? Can it correct factual mistakes, and can it recognize when it does not know an answer?
Researchers also need comparable latency and hardware measurements. A demo recorded in ideal conditions does not establish the experience of a busy restaurant, moving car or crowded call center. Testing should report response delay, interruption accuracy, error rates, energy use and performance over time. Speech naturalness can be evaluated by listeners, but subjective preference should be considered alongside objective task completion.
Finally, developers need deployment details: supported platforms, licenses, integration interfaces, model updates, safety controls and pricing. Until these exist, Dessa voice AI is a promising research announcement rather than a product teams can responsibly schedule for immediate production use.
What Developers and Businesses Should Do Now
For developers interested in real-time voice interfaces, the announcement is a useful prompt to rethink evaluation criteria. Instead of asking only whether an assistant can transcribe and answer accurately, test whether it handles interruptions, preserves context, responds appropriately to emotional cues and remains understandable in noisy environments.
Build a realistic test set before selecting a provider. Include people who speak at different speeds, correct themselves, change topics, interrupt mid-sentence and use different accents. Include failures and recovery: a strong agent should ask for clarification rather than invent an answer when it cannot hear or understand something.
Businesses should avoid committing to a specific unreleased model. Compare available products against their actual requirements and monitor Block’s official documentation for a technical paper, independent evaluations and access terms. A convincing recording is evidence of a possible direction for voice AI, not a substitute for a procurement test.
Frequently Asked Questions
What is Block’s Dessa model?
Dessa is a research-stage speech-to-speech AI model that Block says was trained entirely on recorded human conversations. It can listen and speak simultaneously and is designed to produce more natural conversational timing and expression.
Can I download Dessa now?
No general public download, API or open model weights have been confirmed as of October 26, 2026. Block has described limited early access for AI safety researchers.
Is Dessa open source?
Not based on the published announcement. No public weights or license have been released. A research demonstration should not be described as an open-source launch.
Does Dessa need an internet connection?
Block says the model can run locally on an M4 Max MacBook Pro or better. Whether a future product works entirely offline will depend on its implementation and any connected services.
Is it better than other AI voice assistants?
That has not been established. Block’s demonstrations suggest expressive speech, but independent, standardized comparisons of quality, reliability, latency and safety are not yet available.
The Bottom Line
Dessa voice AI is interesting because it challenges the assumption that advanced conversational systems must begin with text. Block’s speech-only training experiment and full-duplex demonstrations point toward assistants that handle the rhythm of conversation more naturally, potentially with local execution on powerful consumer hardware.
The limits are just as important as the promise. Dessa is not publicly available, its technical details remain incomplete, and selected recordings cannot establish real-world reliability. The next meaningful milestones will be independent safety findings, reproducible benchmarks and clear developer access. Until then, the most useful lesson is broader: future voice assistants should be judged not only by how human they sound, but by how accurately, safely and transparently they help people accomplish tasks.
Ingrid Maldine is a business writer, editor and management consultant with extensive experience writing and consulting for both start-ups and long established companies. She has ten years management and leadership experience gained at BSkyB in London and Viva Travel Guides in Quito, Ecuador, giving her a depth of insight into innovation in international business. With an MBA from the University of Hull and many years of experience running her own business consultancy, Ingrid’s background allows her to connect with a diverse range of clients, including cutting edge technology and web-based start-ups but also multinationals in need of assistance. Ingrid has played a defining role in shaping organizational strategy for a wide range of different organizations, including for-profit, NGOs and charities. Ingrid has also served on the Board of Directors for the South American Explorers Club in Quito, Ecuador.
