🔍 Read the full analysis: The Voice Sounds Human. What Block’s Dessa Can—and Can’t—Show on ThorstenMeyerAI.com
Get the little things that make your day delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Block announced Dessa, a speech-to-speech research model whose company-selected demos feature natural-sounding conversation, laughter and interruptions. The model has not been independently tested or publicly released, so its capabilities and safeguards remain unverified.
Block announced Dessa, a speech-to-speech research model, on October 23, showing recordings in which it laughs, responds to emotion and speaks over a person in conversation. The company says the model was trained from scratch on recorded human conversation, but no independent researcher or reporter has tested it directly, and Block has not released a paper, benchmarks, model weights or an API.
Block, the company behind Square and Cash App, describes Dessa as a model that listens and speaks at the same time rather than converting speech to text before generating a reply. Its project page says it was trained using human conversation only, without text or pretrained models, and can recognize features such as laughter, sarcasm and changes in emotional tone. Those are company claims, not independently verified results.
The recordings selected by Block include laughter during a sentence, a conversation in which the model helps retrieve a forgotten film title, and a response to news of an anniversary surprise. Block says the clips are live conversations that were not steered. Co-founder Jack Dorsey wrote that speaking with Dessa feels like talking with a person, citing its emotion, pauses and cadence. The demos do not establish how the system performs in other conversations or whether the same results can be reproduced.
Block also says Dessa runs faster than real time on a MacBook Pro M4 Max or better and was trained using one NVIDIA GB200 NVL72. It has not published a latency measurement, model size, benchmark results or a public release date. Access is limited to a waitlist and early access for a small group of AI safety researchers, according to the source material.
Outside coverage has offered a more qualified reading of the demonstrations. The Verge described some exchanges as awkward and uncanny while finding Dessa more natural than another model in the comparison. TechCrunch said it was unclear whether emotional qualities were being guided by tags. The Decoder cautioned that sounding emotional does not show that a system feels emotion or has human-like understanding.
Infographic 01 · Dessa fact sheet
What Block claims, what was heard, what is missing.
Every capability below comes from Block or its selected recordings. Nobody outside Block has run the model yet.
| Topic | What we know | Status |
|---|---|---|
| Training data | Human conversation only: “no text, no pretrained models” | Block’s claim |
| Conversation | Full-duplex: listens while it talks, backchannels, cuts in | Block’s claim |
| Emotion | Hears and speaks laughter, sarcasm, sighs; unclear if steered by tags (TechCrunch) | Block’s claim |
| Speed | “Faster than real time”; no latency figure | Block’s claim |
| Hardware | Trained on one NVIDIA GB200 NVL72; runs on a MacBook Pro M4 Max or better | Block’s claim |
| Demos | Some chats “a little awkward and uncanny”, more natural than GPT-Live-1 (The Verge) | reporters heard it |
| Size, benchmarks, release date | Not disclosed (CoinDesk) | not disclosed |
| Paper, model card, license, weights | None found as of 26 Oct 2026 | not disclosed |
| Languages | Not listed; Block calls learning new languages an open question | not disclosed |
| Voice cloning safeguards | No statement on watermarking, consent checks or imitating a specific person | not disclosed |
Who can use it
research preview
- Public: a waitlist; “launch links are coming soon”
- Early access: “a small group of AI safety researchers”, institutional email required
- Business use: none announced; Square already uses voice AI for ordering (TechCrunch)
Who built it
from Block’s page
- A “(very) small AI research lab at Block” reporting to Jack Dorsey
- Founded by former members of Dessa, the Toronto startup Square acquired in 2020 (CoinDesk)
- Lead Ragavan Thurairatnam: “a GPT3 moment for speech”
Sources: Block, Dessa project page (23 Oct 2026); Ragavan Thurairatnam on X (23 Oct 2026); The Verge (23 Oct 2026); TechCrunch (24 Oct 2026); CoinDesk (24 Oct 2026).
Powered by Thorsten Meyer AI
When Natural Voices Meet Fraud Risks
Dessa matters beyond the performance of one research model because convincing voice impersonation is already a fraud concern. The FBI’s 2025 Internet Crime Report, released in April 2026, counted 22,364 complaints mentioning AI and $893 million in reported losses. The report also recorded more than $5 million in losses from distress scams involving cloned voices of loved ones. Those figures cover reported cases; they do not show that Dessa was used in any crime.
A more natural-sounding system could make it harder for listeners to rely on pauses, awkward replies or flat delivery as signs that a caller is artificial. But the Block demos do not establish that Dessa can imitate a particular person, and the company has not said whether the model includes safeguards against doing so. The distinction matters: the potential risk is broader than this unreleased model, while claims about Dessa’s specific misuse remain untested.
For organizations, the announcement is a reminder that a familiar-sounding voice alone is not strong proof of identity. A safer process for financial requests is to verify them through a separate, known contact channel. For families, a code word agreed in advance can help check an urgent call. These precautions address existing voice-cloning risks; the source does not report that Block has issued them as formal guidance.
As an affiliate, we earn on qualifying purchases.
What Block Has and Has Not Released
Block published its project page on October 23, 2026. The research effort is described as a small lab at Block, founded by former members of the Toronto deep-learning startup Dessa, which Square acquired in 2020, according to CoinDesk. Lead Ragavan Thurairatnam called the work “a GPT-3 moment for speech” and said it could run on consumer GPUs. That characterization is his assessment, not an independent benchmark finding.
The public evidence so far consists of Block’s selected recordings and descriptions, alongside reporting by outlets that listened to the clips. There is no public model, technical paper, model card or license identified in the supplied material. Without those materials, outsiders cannot check the training details, reproduce the demos, measure performance or evaluate the system’s protections.
Block’s stated design also points to an unresolved question about how conversational behavior is produced. The company says the model can speak and listen simultaneously and respond to emotional cues, while reporters have raised questions about possible steering. The available demos show the resulting audio, but do not disclose enough to settle how prompts, tags or other controls affect its voice.
“Talking with it feels like talking with a person… full of emotion, nuance, pauses, and a natural cadence.”
— Jack Dorsey, Block co-founder
Performance and Protections Remain Untested
Dessa’s real-world performance is unknown. No outside party has run the model, and the supplied material identifies no paper, model card, benchmarks, public release date, license or API. Block has not published the model’s size or a latency figure, despite saying it runs faster than real time.
Other open questions include which languages it supports, whether emotional delivery is directed by tags, and what rules or technical safeguards govern voice imitation. The source material says Block has not explained whether it uses watermarking or consent checks, or whether it can imitate a specific person. Its claim that Dessa chooses its own voice and personality does not answer those questions.
The FBI figures provide a measure of reported AI-related fraud and selected scam losses, but they do not establish how common cloned-voice scams are overall or connect any reported loss to Dessa. No evidence is presented that the model has been used in fraud.
Independent Access Is the Next Test
The next meaningful evidence would be access for independent researchers, paired with documentation that describes the model, its evaluation and its safeguards. Block has said a small group of AI safety researchers has early access, but the source material does not give a schedule for broader testing or release. Until results from outside Block are available, the demos should be treated as illustrations of the company’s claims, not proof of general capability.
Companies and consumers do not need to wait for a public Dessa release to review voice-verification practices. Organizations can require separate confirmation for payment changes and urgent transfer requests. Individuals can end a suspicious call and use a number they already trust to reach the person or institution. These steps address current impersonation risks without assuming that a particular new model is responsible.
Key Questions
What is Block’s Dessa?
Dessa is a speech-to-speech research model that Block says can listen and speak at the same time. The company’s demos show conversational exchanges, but the model is not publicly available for independent testing.
Has anyone outside Block verified its capabilities?
Not by running the model, according to the source material. Reporters have listened to Block’s recordings, but there is no public paper, benchmark or independent test of the system itself.
Can Dessa clone a specific person’s voice?
That has not been established. The source material says Block has not explained whether Dessa can imitate a specific person or what consent checks may apply.
Is there evidence that Dessa has been used in scams?
No evidence is provided that Dessa has been used in fraud. The FBI’s reported AI-related complaints and losses describe broader fraud concerns, not cases linked to this model.
How can people check an urgent voice request?
End the call and contact the person or organization using a number already known to you. Families can also agree on a code word in advance; companies can require separate verification for payment requests.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
