A Voice That Laughs Isn’t Proof: Block’s Dessa And AI Speech

🔍 Read the full analysis: A Voice That Laughs Isn’t Proof: Block’s Dessa And AI Speech on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

• 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 has introduced Dessa, a speech-to-speech research model whose company-published demos show it laughing, interrupting and responding with emotional-sounding speech. The announcement matters because natural, responsive voices could make spoken AI easier to use—and harder to distinguish from a person—but no outside researcher has tested the model, and Block has not published a technical paper, benchmarks or public access.

Block, the company behind Square and Cash App, published a project page for Dessa on October 23. The company says it trained the model from scratch on recorded human conversation, without text or pretrained models. It describes the system as full-duplex: Dessa can listen while speaking, respond to interruptions and provide conversational signals such as laughter or backchanneling.

Block also says inference runs faster than real time on a MacBook Pro M4 Max or better, and that training used one NVIDIA GB200 NVL72. It has not supplied a latency measurement, model size or benchmark results. Those specifications are company statements, not independently verified performance findings.

The recordings on Block’s page include a conversation in which Dessa laughs mid-sentence, a prompt about a forgotten film title and a response to news of an anniversary surprise. Block says the clips are live recordings and that nobody steered the model. They remain Block-selected demonstrations. The Verge found some exchanges awkward or uncanny while describing Dessa as more natural than another model in the comparison; TechCrunch noted that it was unclear whether emotion was guided by tags.

Natural Voices Complicate Caller Checks

Dessa’s demos put a familiar trade-off in sharper focus: voice AI that responds with timing and emotional cues may feel more comfortable in ordinary services, but those same cues can make a synthetic caller more persuasive. A convincing laugh or interruption is not evidence that the speaker is a real person, and a natural-sounding voice is not a reliable identity check.

The risk exists beyond this one unreleased model. The FBI’s 2025 Internet Crime Report, published in April 2026, counted 22,364 complaints mentioning AI and $893 million in reported losses. It separately reported more than $5 million in losses tied to distress scams using cloned voices of loved ones. These figures concern reported cases, not a measure of Dessa’s capabilities or proof that the model has been used in scams.

For organizations, the practical implication is that voice alone should not authorize payments or sensitive account changes. Independent verification—such as calling back using a number already on file—can help avoid relying on caller ID or conversational realism. For families, agreeing on a code word and treating unexpected urgency as a reason to pause are low-cost precautions. These measures address voice impersonation generally; the available material does not show that Dessa itself imitates specific people.

From Block’s Research Lab to Demos

Block describes the team behind Dessa as a small research lab founded by former members of the Toronto deep-learning startup Dessa, which Square acquired in 2020, according to CoinDesk. The project is presented as research, not as a released consumer product. Block has offered a waitlist and says early access is limited to a small group of AI safety researchers.

Block co-founder Jack Dorsey promoted the model’s conversational quality, while project lead Ragavan Thurairatnam compared the work to a major moment for speech technology. Those are the team’s assessments, not independent conclusions about general capability. The published demonstrations let viewers hear selected exchanges, but do not establish how consistently the model handles unfamiliar prompts, noisy conditions or conversations outside the examples.

The fraud concern predates Dessa. The FBI report describes voice cloning in distress scams and notes AI-assisted business email compromise, including situations where cloned voices can be used to request wire payments. The report’s figures provide context for the broader risk; they do not show that Block’s model contributed to any reported loss.

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

— Block co-founder Jack Dorsey

Testing, Access and Safeguards Remain Open

As of October 26, Block had not released a research paper, model card, benchmark results, license or public model access. The model’s size, full technical design and a release date were also undisclosed. The company’s performance and training statements therefore cannot yet be checked against independent evaluation.

Other open questions include which languages Dessa supports, how it performs beyond curated recordings, and whether its emotional responses are generated directly or influenced by controls such as tags. Block says the model chooses its own voice and personality, but the material reviewed does not establish whether it can imitate a particular person. It also does not specify watermarking, consent checks or other safeguards for preventing misuse. No evidence in the supplied material shows that Dessa has been used to defraud anyone.

Independent Access Will Set the Test

The next meaningful development will be whether Block gives independent researchers access and publishes enough technical detail to test its claims. Evaluation could clarify conversational reliability, performance across languages and conditions, and how the model handles identity, consent and safety concerns. Block has said a small group of AI safety researchers will receive early access, but it has not announced a date for broader release.

Until more evidence is available, readers should treat the recordings as demonstrations rather than proof of general performance. Businesses can review whether their payment and account-change procedures rely too heavily on recognizing a voice. Families can agree on a verification phrase and call back using a trusted number if a caller creates pressure. The broader question—how to label and safeguard increasingly realistic synthetic speech—will depend on evidence about deployed systems, not demonstrations alone.

Key Questions

What is Block’s Dessa?

Dessa is a speech-to-speech research model that Block says was trained on recorded human conversation. The company says it can listen and speak at the same time, including responding to interruptions and conversational cues.

Has anyone outside Block tested the model?

Not according to the information available as of October 26, 2026. Reporters have listened to Block’s recordings, but no independent hands-on test or public benchmark results had been reported.

Can Dessa imitate a specific person?

The available material does not establish that. Block says Dessa picks its own voice and personality, but has not disclosed details about person-specific voice imitation or related consent safeguards.

Does the FBI data show that Dessa is being used in scams?

No. The FBI’s 2025 figures describe complaints and losses involving AI and voice-cloning scams generally. They do not link those cases to Dessa.

How can people respond to a suspicious urgent call?

Do not rely on a familiar-sounding voice alone. Hang up and call the person or organization using a number you already trust; families can also agree on a private code word for urgent requests.

Source: ThorstenMeyerAI.com