Can A Human-Sounding AI Voice Tell Us Who—or What—is Speaking?

Read the full analysis: Can A Human-Sounding AI Voice Tell Us Who—or What—is Speaking? on ThorstenMeyerAI.com

Block announced Dessa, a speech-to-speech research model, on October 23, 2026, sharing recordings in which it laughs, responds to emotion and talks over or alongside a human speaker. The demos suggest a more conversational style of voice AI, but no outside researcher has tested the model, and Block has not published a paper, benchmarks, model weights or an API. The announcement has also sharpened concerns that more natural-sounding AI could make voice impersonation harder to identify.

Block says Dessa was trained from scratch on recorded human conversation, without text or pretrained models. The company describes it as full-duplex: it can listen while speaking, interrupt and use conversational backchannels. Block also says it can convey qualities such as sarcasm, laughter and sadness, and that inference runs faster than real time on a MacBook Pro M4 Max or better. It has not supplied a specific latency figure or independent evidence for those claims.

The recordings on Block’s project page include a laugh mid-sentence, help recalling a film title and a response to an anniversary surprise. Block says the clips are live conversations and that nobody steered the model. Co-founder Jack Dorsey wrote that talking with it feels like talking with a person, citing emotion, nuance, pauses and cadence. Dessa lead Ragavan Thurairatnam called it “a GPT3 moment for speech.” These are company and project-team assessments, not independent test results.

Outside coverage has been more qualified. The Verge described some exchanges as occasionally “a little awkward and uncanny,” while finding Dessa more natural than another model used in a comparison. TechCrunch said it was unclear whether emotional qualities were being steered by tags. As of October 26, Block had not disclosed the model’s size, benchmark results, language coverage or a public release date. Access is limited to a waitlist and a small group of AI safety researchers, according to the source material.

Natural Voices and Fraud Risks

The concern is not that Dessa has been shown to commit fraud; there is no evidence of that in the available material. The issue is that conversational cues such as pauses, emotional responses and interruptions can make synthetic speech feel more like a live exchange. If voice interactions become harder to distinguish from human ones, familiar checks—such as trusting a caller because they sound like someone you know—may be less dependable.

The FBI’s 2025 Internet Crime Report, published 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 describe reported crimes, not losses caused by Dessa, and the source does not establish that its model has been used in scams.

For organizations, the development is a reason to avoid treating voice alone as proof of identity. The source recommends callbacks for payment requests and clear disclosure when a voice is synthetic. For individuals, a pre-agreed family code word and calling back on a known number can add a separate check. These precautions address existing impersonation risks; they are not evidence that Dessa itself is publicly available or being misused.

What Block Has Released So Far

Block, the company behind Square and Cash App, published a project page for Dessa on October 23. The team is described as a small research lab at Block, founded by former members of Toronto-based deep-learning company Dessa, which Square acquired in 2020. Block says the model was trained using one NVIDIA GB200 NVL72 and learned from far fewer words than large language models typically require. Those details come from the company; a technical paper or reproducible evaluation has not been published.

The demos provide the main public evidence available so far. They can show what selected recordings sound like, but they do not establish how the model performs across speakers, languages, noisy conditions or longer conversations. Nor do they answer whether the model can imitate a specific person. Block says Dessa chooses its own voice and personality, but the available information does not include a detailed account of voice controls or consent safeguards.

The source also points to a regulatory and practical setting: it says the EU AI Act’s Article 50 applies from August 2, 2026, including transparency obligations relevant to synthetic content. The precise obligations depend on the use and circumstances. The announcement does not by itself show how Block will meet any applicable requirements.

“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 Safeguards Unverified

Independent performance remains untested. There is no published research paper, model card, benchmark suite, public model release or API in the source material. Block has not given a public release date or a list of supported languages. The demos do not establish how consistently the model handles interruption, emotion or unfamiliar conversation outside the examples selected for publication.

Other open questions include whether emotional delivery can be controlled with tags, what consent checks or watermarking may be used, and whether the system can reproduce a particular person’s voice. Block’s statement that Dessa selects its own voice and personality does not answer those technical or policy questions. It is also unknown whether the model’s local hardware requirements or access limits will change.

Awaiting Independent Model Access

The next useful evidence would be a technical paper, model card, benchmarks and access that allows researchers to test the system independently. Block has said a small group of AI safety researchers has early access, but the source provides no timetable for broader availability. Until those details emerge, claims about its capabilities should be treated as Block’s account of a research model, rather than established results.

For now, the practical response does not depend on Dessa becoming public: organizations can verify sensitive requests through a separate channel, and families can agree on a code word and call back using a number they already trust. Whether Block adds specific disclosure, consent or anti-impersonation measures—and how the model performs beyond its launch recordings—remains to be seen.

Key Questions

What is Block’s Dessa?

Dessa is a speech-to-speech research model that Block says was trained from scratch on recorded human conversation. Block says it can listen while speaking and handle conversational features such as interruptions, laughter and emotional expression.

Has an independent researcher verified the demos?

Not according to the available source material. The public evidence is Block-selected recordings; no outsider has run the model, and no paper, benchmark results or model weights have been published.

Does the announcement show that Dessa has been used for fraud?

No. The source does not report Dessa being used in a scam. The concern is that more natural-sounding speech could complicate voice-based identity checks, against a backdrop of existing voice-cloning fraud.

How can people check a suspicious call?

Do not rely on the caller’s voice alone for a sensitive request. Hang up and call a known number, verify payment requests through a separate channel, and consider agreeing on a family code word for emergencies.

When will the model be available to the public?

Block had announced no public release date as of October 26, 2026. The source says access is limited to a waitlist and a small group of AI safety researchers.

Source: ThorstenMeyerAI.com