🔍 Read the full analysis: What Happens When AI Speech Includes Laughter And Interruptions? on ThorstenMeyerAI.com
Get the latest gadgets 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 that it says can handle laughter, emotion and interruptions. The demonstrations are company-selected, and the model has not been made available for independent testing; its realism also raises concerns about voice impersonation and fraud.
Block announced Dessa, a speech-to-speech research model, on October 23, with company recordings that show it laughing mid-sentence, responding to emotion and interrupting naturally. The examples point to a potential change in how voice assistants converse, but the model has not been independently tested or publicly released, leaving its performance and its risks in real-world calls unverified.
Block says Dessa was trained from scratch on recorded human conversation alone, without text or pretrained models. The company describes it as full-duplex: it listens while speaking and can decide when to respond, interrupt or give brief conversational feedback. Block also says the system recognizes and produces cues such as sarcasm, laughter and frustrated sighs, which can be lost when speech is converted into text before processing.
The recordings on Block’s project page include a laugh during a sentence, a conversation about forgetting a film title and an emotional response to an anniversary surprise. Block says each clip is a live recording and that nobody steered the conversation. The examples are Block-selected demonstrations, however, not results from an independent evaluation. The Verge described some exchanges as “a little awkward and uncanny,” while finding Dessa more natural-sounding than another model included in the demonstrations.
Block says the model runs faster than real time on a MacBook Pro M4 Max or better, but has not published a latency figure. The company also says training used one NVIDIA GB200 NVL72. It has not released the model’s size, benchmarks, technical paper, model card, license or public access; the project page offers a waitlist and says a small group of AI safety researchers has early access.
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
Natural Speech Raises Fraud Risks
Conversational details matter because they can make a voice system feel more responsive and less scripted. Laughter, pauses, backchannels and interruptions may improve accessibility or make customer-service exchanges easier. But the same features can make an impersonation more convincing to someone who expects a real person on the other end. Natural-sounding speech is not proof of identity, and a familiar voice should not be treated as a secure check for approving a payment or sharing sensitive information.
The FBI’s 2025 Internet Crime Report, published in April 2026, recorded 22,364 complaints mentioning AI and $893 million in reported losses. The report said victims lost more than $5 million to distress scams using voice cloning to mimic a loved one. Those figures describe reported complaints and losses across the report’s categories; they do not show that Dessa was used in fraud or quantify the effect of this model. The risk is that more convincing speech could add to an existing problem.
For organizations, the development reinforces the case for checking requests through a separate, previously verified channel instead of relying on caller recognition. For families, an agreed code word and a call back to a known number can help verify an urgent request. Those are practical safeguards, not guarantees, and the demonstrations alone do not establish how easily Dessa or other systems can imitate a particular person.
Products Worth Considering
LG UltraGear 27-Inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDMI 2.1
Logitech MX Master 4 for Mac, Ergonomic Bluetooth Mouse with Haptic Feedback, Ultra-Fast Scrolling, USB-C Charging, for MacOS, iPadOS, Windo
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
As an affiliate, we earn on qualifying purchases.
From Text Pipelines to Live Conversation
Many voice assistants process speech by turning it into text, generating a response and then converting that response back into audio. Block argues that this approach discards emotional and conversational signals. Dessa is presented as a model that works directly with speech and can listen and speak at the same time, rather than waiting for a complete turn before responding. That distinction is the central technical claim behind the demonstrations.
Block, the company behind Square and Cash App, said a small research group developed the project. Co-founder Jack Dorsey described talking with Dessa as feeling like talking with a person, citing emotion, nuance, pauses and cadence. Project lead Ragavan Thurairatnam called it “a GPT-3 moment for speech” and said it could run on consumer GPUs. These are company and team assessments, not conclusions established by published benchmarks or external testing.
Outside observers have raised questions about how the voice is generated and controlled. TechCrunch reported that it was unclear whether emotional expression might be steered by tags. The Decoder cautioned that sounding emotional is not the same as being emotional or intelligent. Block says Dessa selects its own voice and personality, but that does not answer whether it can imitate a specific person or what safeguards govern voice selection.
“Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.”
— Jack Dorsey, Block co-founder
Testing and Safeguards Still Missing
Independent access is limited, so outside researchers and reporters have not been able to test how consistently the model handles interruptions, emotion, noisy audio or unusual conversations. Block has not published benchmark results, a model card or a technical paper. Without those materials, it is not possible to compare its performance with other systems or confirm the company’s claims about speed, training efficiency and conversational ability.
Several safety questions also remain unanswered. Block has not publicly detailed its approach to watermarking, consent checks or preventing a user from imitating a specific person. The company’s statement that the model picks its own voice and personality does not settle whether a user could prompt it to resemble someone else. The list of supported languages and the timing of any broader release are also unknown. The recordings do not establish whether the model has been used in scams or whether it can reliably evade detection.
Awaiting Independent Access
The next useful evidence would be access for independent researchers, alongside a technical paper or model documentation describing the system’s limits, safeguards and evaluation methods. Block has said that a small group of AI safety researchers has early access, but it has not announced a public release date, API or release of model weights. No independent test results were available as of October 26, 2026.
Until those details emerge, the demonstrations should be treated as evidence of what Block chose to show, not as proof that the model performs consistently in ordinary calls. Companies and households can act on the broader lesson now: verify high-stakes requests through a known, separate channel and do not rely on a familiar-sounding voice alone.
Key Questions
What is Block’s Dessa?
Dessa is a speech-to-speech research model that Block says can listen and talk at the same time and respond to conversational cues such as laughter and pauses.
Has Dessa been independently tested?
Not publicly as of October 26, 2026. Block’s examples are company-selected recordings, and the company has not released a paper, model card, benchmarks or general public access.
Does the announcement show that Dessa has been used for fraud?
No. The announcement and demonstrations do not show that Dessa has been used in scams. The concern is that more natural voice technology could compound existing voice-impersonation risks.
How can people check an urgent request made by phone?
End the call and contact the person using a number already saved or independently verified. Families may also agree on a code word. A convincing voice alone should not authorize a payment or disclosure of sensitive information.
Source: ThorstenMeyerAI.com
