How To Think About AI Voices That Laugh, Hesitate, And Interrupt

🔍 Read the full analysis: How To Think About AI Voices That Laugh, Hesitate, And Interrupt on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

Block announced Dessa, a speech-to-speech research model that the company says can laugh, pause, respond to emotion and interrupt naturally, and published recordings of it on October 23. The demos offer a glimpse of more humanlike voice AI, but the model has not been independently tested or publicly released, leaving its capabilities and implications for voice fraud 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 can listen while speaking and decide when to respond or provide a conversational backchannel. Its project page also says it can recognize and produce cues such as sarcasm, laughter and a frustrated sigh. Those are company claims; Block has not published a technical paper, model card, benchmark results or model weights.

The company says the system runs faster than real time on a MacBook Pro M4 Max or better, but has not provided a latency figure. Block also says it trained the model using one NVIDIA GB200 NVL72. Co-founder Jack Dorsey wrote that talking with it feels like talking with a person, while Dessa lead Ragavan Thurairatnam called the work “a GPT3 moment for speech.” Neither statement is an independent performance assessment.

Block’s recordings include Dessa laughing mid-sentence, helping a speaker recall a film title and responding to an anniversary surprise. Block says the clips are live conversations that were not steered. The Verge found the exchanges more natural than a comparison model in some respects, but also “a little awkward and uncanny.” TechCrunch reported that it was unclear whether emotion in the demos might be guided by tags. Only Block’s selected recordings are available for outside assessment; access is currently limited to a waitlist and a small group of AI safety researchers, according to the source material.

Natural Voices Complicate Fraud Checks

Dessa’s reported conversational cues matter because people often use vocal behavior to judge whether a caller sounds familiar or genuine. A voice that can pause, laugh or respond to interruptions could make it harder to rely on conversational awkwardness as a warning sign. That does not show that Dessa has been used for fraud, or that it can clone a specific person. It does highlight why a familiar-sounding voice should not be treated as proof of identity.

The FBI’s 2025 Internet Crime Report, published in April 2026, counted 22,364 complaints mentioning AI and reported losses of $893 million. The FBI also said victims lost more than $5 million to distress scams involving cloned voices of loved ones. These are figures for the report’s stated categories and period; they do not establish that Dessa caused, or was involved in, any of those losses.

For organizations, the practical response is to verify consequential requests through a separate, trusted channel rather than voice alone. For families, agreeing on a private code word and calling back using a known number can provide a check when a caller claims an emergency. Urgency should prompt verification, not an immediate transfer of money or disclosure of sensitive information.

From Speech Research to Public Demos

Block, the company behind Square and Cash App, published the Dessa project page on October 23, 2026. Its team is described as a small research lab formed by former members of the Toronto AI startup Dessa, which Square acquired in 2020, according to CoinDesk. The announcement positions the project as research into speech that retains conversational and emotional information, rather than a conventional system that first converts speech to text and then generates a reply.

The project is not presented as a broadly available product. Block has announced no public API, model weights or release date. The evidence available so far consists of Block’s own recordings and the limited reporting by outlets that listened to them; journalists have not run the model themselves. That distinction matters: a polished demonstration can show examples of a system’s output, but cannot establish its reliability across callers, situations or languages.

Block’s own page leaves the model’s language-learning efficiency as an open question. Other unresolved technical details include its size and how its emotional behavior is produced. The Decoder cautioned that sounding emotional is not the same as being emotional or intelligent. Humanlike delivery is an observable presentation, not evidence of human feelings or understanding.

“Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.”

— Jack Dorsey, Block co-founder

Capabilities Still Lack Outside Tests

As of October 26, 2026, there is no public technical paper, model card, benchmark suite, license or release date in the source material. No outside researcher or journalist has reported running Dessa, so its performance beyond the company’s recordings remains unverified. Block has not disclosed the model’s size or a numerical latency result, and the range of languages it supports is unclear.

It is also not clear whether the emotional effects in the recordings are generated entirely by the model or partly guided by tags. Block says Dessa develops its own voice and personality, but the available material does not establish whether it can imitate a particular person, what safeguards or consent checks it uses, or whether outputs are watermarked. There is no evidence in the supplied reporting that Dessa has been deployed in scams; the concern is about the potential misuse of increasingly natural voice systems in general.

Awaiting Access and Technical Evidence

The next evidence to watch for is whether Block expands access beyond its waitlist and safety-research group, publishes technical documentation or provides independent evaluators with a way to test the model. Those steps could help clarify its speed, reliability, supported languages and safeguards. Block has not announced a timetable, so it is not yet clear when any wider evaluation or release might occur.

Until then, consumers and businesses can apply safeguards that do not depend on identifying a synthetic voice by ear: independently confirm payment instructions, call alleged relatives back on a number already known, and agree on a private verification phrase for emergencies. In the European Union, the source material says AI Act Article 50 applies from August 2, 2026; organizations using synthetic voices should check the disclosure obligations relevant to their use. The precise duties may depend on the system and circumstances.

Key Questions

What is Block’s Dessa?

Dessa is a speech-to-speech research model announced by Block on October 23, 2026. Block says it was trained on recorded human conversation and can respond with cues such as laughter, pauses and interruptions.

Can the public use or test Dessa?

Not broadly, based on information available as of October 26, 2026. Block has announced no public model weights or API; it says access is limited to a waitlist and a small group of AI safety researchers.

Has Dessa been shown to clone a specific person’s voice?

The available reporting does not establish that. Block says the model picks its own voice and personality, but it has not supplied enough information to determine whether it can imitate a particular person or what consent protections apply.

How can people respond to a suspicious voice call?

Do not treat a familiar voice as authentication for an urgent request. Hang up and call back using a number you already trust, verify payment instructions through another channel, and consider agreeing on a family code word for emergencies.

Source: ThorstenMeyerAI.com