Block’s Dessa Sounds Natural. Does Natural Speech Mean Human?

🔍 Read the full analysis: Block’s Dessa Sounds Natural. Does Natural Speech Mean Human? on ThorstenMeyerAI.com

Age 18–24?Offer from Amazon

Prime made for students and young adults

• Fast, free delivery for dorm and study essentials
• Prime Video and Amazon Music included
• Member-only deals

Try Prime for Young Adults →Free trial for eligible 18–24 year olds

As an affiliate, we earn on qualifying purchases.Hide

Block has demonstrated Dessa, a research model that talks and listens at the same time and, in company recordings, laughs, pauses and responds to interruptions. The October 23 announcement has drawn attention both for the model’s conversational style and for the possibility that more lifelike AI voices could make impersonation scams harder to detect; independent testing has not been reported.

Block says it trained Dessa from scratch on recorded human conversation, without text or pretrained models. The company describes the system as full-duplex: it can listen while speaking, interrupt and provide conversational backchannels. Its project page also says the model can handle vocal cues such as sarcasm, laughter and sighs, and that inference runs faster than real time on a MacBook Pro M4 Max or better. Block has not provided a specific latency figure.

The demonstrations on Block’s project page include laughter during a sentence, a conversation in which the speaker forgets a film title, and a response to an anniversary surprise. These are Block-selected recordings. The company says each is a live conversation and that nobody steered the model, but outsiders have not been given general access to reproduce the results. The Verge described some exchanges as “a little awkward and uncanny,” while saying Dessa came across as more natural than another model featured in a comparison.

Block has announced no public model weights or API. As of October 26, there is no published technical paper, model card, benchmark report, license or release date in the supplied reporting. Block says a small group of AI safety researchers has early access and others can join a waitlist. That means the model’s performance, limitations and suitability for use outside the demonstrations remain unverified.

Natural Voices and Fraud Risks

The concern is not that Dessa has been shown to commit fraud; the demonstrations do not establish that. The issue is that conversational realism can make it harder for a listener to rely on pauses, flat delivery or awkward timing as signs that a voice is synthetic. The same qualities that could make voice interfaces easier to use may also help an impersonator sound more credible.

The source material cites the FBI’s 2025 Internet Crime Report, released in April 2026, which recorded 22,364 complaints mentioning AI and $893 million in reported losses. It also says victims lost more than $5 million to distress scams involving cloned voices of loved ones. These figures describe reported complaints and losses across AI-related crime; they do not show that Dessa was involved or measure the effect of this model.

For readers, the practical lesson is to avoid treating a familiar voice as proof of identity when money or urgent action is involved. The source recommends calling back using a number already known to the recipient and agreeing on a family code word. For organizations, voice alone is a weak basis for approving payments or changing account details; a separate verification channel can reduce reliance on a convincing call.

What Block Has Demonstrated

Block, the company behind Square and Cash App, published a project page for Dessa on October 23, 2026. Co-founder Jack Dorsey described conversations with it as feeling like talking with a person, citing emotion, nuance, pauses and natural cadence. Dessa lead Ragavan Thurairatnam called the project “a GPT3 moment for speech,” while Block’s announcement emphasizes the model’s speech-only training and ability to handle vocal expression.

Those statements are company claims, not results from a published independent evaluation. TechCrunch reported that it was unclear whether emotional expression might be steered by tags. CoinDesk noted that Block had not disclosed the model’s size, benchmark results or a public release date. The Decoder cautioned that sounding emotional does not establish that a system is emotional or intelligent. The distinctions matter because a polished demonstration can show what a system can do in selected situations, but not how reliably it performs across callers, languages or difficult conversations.

The source also reports that Block says Dessa was trained using one NVIDIA GB200 NVL72 and learned from far fewer words than large language models. Without a technical paper or detailed training documentation, those descriptions cannot be independently checked. Block calls learning new languages an open question, and its project page says the model picks its own voice and personality. The available material does not establish whether it can imitate a specific person.

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

— Jack Dorsey, Block co-founder

What the Demos Cannot Verify

There is not yet enough public information to establish how Dessa performs beyond Block’s recordings. No independent evaluation, benchmark results or technical paper are available in the source material, and the model is not publicly accessible. Its reliability across accents, languages, noisy conditions and longer conversations is unknown.

Other open questions include how the system handles consent, whether it has safeguards against impersonation, and whether its output is watermarked or otherwise identifiable as synthetic. Block has not announced a public release date or said whether people will be able to use the model to reproduce a particular person’s voice. The company’s claim that Dessa chooses its own voice does not resolve those questions.

The fraud figures provide context about existing reported harms, but they do not establish that a more natural model will cause a particular increase in scams. No evidence in the supplied material links Dessa to any fraud. The consequences of its release, if it becomes broadly available, will depend on access, safeguards and how people and organizations use it.

Access and Independent Testing

The next useful evidence would be a technical paper or model card, followed by testing that can be replicated by researchers outside Block. Such material could clarify the training setup, supported languages, measured latency, performance limits and safeguards. Block has not provided a date for those disclosures or for a wider release.

For now, access is limited to a waitlist and early access for a small group of AI safety researchers, according to the source. Readers should treat the examples as demonstrations rather than proof of general capability. If Block expands access or publishes evaluation results, those details will help show whether Dessa’s conversational qualities hold up beyond its launch recordings and what protections accompany them.

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. The company says it was trained from scratch on recorded human conversation without text or pretrained models.

Has an independent organization tested Dessa?

Not in the reporting supplied here. Publicly available evidence consists of Block-selected recordings; no independent benchmark results or public model access have been reported.

Does Dessa prove that AI can feel emotion?

No. A system’s ability to produce emotional-sounding speech does not establish that it experiences emotion or understands it as a person would. The source material includes that distinction in The Decoder’s assessment.

Is there evidence Dessa has been used in a scam?

No evidence in the supplied reporting links Dessa to a scam. The fraud discussion concerns broader risks from voice cloning and the possibility that natural conversational cues could make synthetic voices harder to recognize.

When will the public be able to use Dessa?

Block has not announced a public release date or an API or model weights. The source says access is currently through a waitlist and a small early-access group of AI safety researchers.

Source: ThorstenMeyerAI.com