A Human-Sounding Voice Isn’t Proof: The Case Of Block’s Dessa

🔍 Read the full analysis: A Human-Sounding Voice Isn’t Proof: The Case Of Block’s Dessa 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 published demonstrations of Dessa, a speech-to-speech research model that the company says can respond in real time with laughter, pauses and interruptions. The clips present a voice that can sound conversational, but no outside researcher has tested the model, and Block has not released technical documentation or public access—leaving its capabilities and risks difficult to verify.

Block announced Dessa on October 23 on a project page. The company says it trained the model from scratch on human conversation alone, without text or pretrained models. It describes the system as full-duplex: it can listen while speaking, interrupt and respond to conversational cues such as laughter or a frustrated sigh. Block also says it runs faster than real time on a MacBook Pro M4 Max or better, but did not provide a specific latency figure.

The company’s selected recordings include a voice laughing mid-sentence, helping a user recall a film title and responding to news of an anniversary surprise with what the source describes as teary gratitude. Block says the recordings are live conversations and that no one steered the model. Those demonstrations are evidence of what Block chose to show, not an independent evaluation of performance across users, situations or languages.

Outside coverage has noted limitations and open questions. The Verge described some exchanges as “a little awkward and uncanny”, while finding Dessa more natural-sounding than another model in one comparison. TechCrunch reported that it was unclear whether emotion in the output was guided by tags. As of October 26, Block had not published a research paper, model card, benchmarks, license or release date. The company says a small group of AI safety researchers has early access; a broader waitlist is available, but there are no announced weights or API.

Why Natural-Sounding Calls Raise Risks

Dessa’s potential importance lies not in proof that it can impersonate a particular person, but in the broader possibility that conversational voice systems may become harder to distinguish from people. Natural timing, interruptions and emotional cues can make a call feel more convincing. That could improve customer service, but those same qualities may also help a fraudster sustain an impersonation. Block has not said whether Dessa can imitate a specific person, and the demonstrations do not establish that it can.

The concern sits alongside existing fraud, not a hypothetical future threat. The FBI’s 2025 Internet Crime Report, published in April 2026, counted 22,364 complaints mentioning AI and reported $893 million in losses associated with those complaints. It also recorded more than $5 million in losses from distress scams using cloned voices of loved ones, according to the source material. These figures cover reported complaints and losses in 2025; they do not measure Dessa or show that its release caused any increase.

For organizations, the practical issue is that a familiar-sounding voice should not stand alone as proof of identity when approving payments or changing account details. Verification through a known callback number and a second channel can reduce reliance on a caller’s voice. For families, a pre-agreed code word and calling back using a saved number are straightforward safeguards. These measures address voice impersonation generally; there is no evidence in the material that Dessa has been used in a scam.

Block’s Speech-Only Model Claims

Block, the company behind Square and Cash App, describes Dessa as the work of a small research lab founded by former members of Toronto deep-learning startup Dessa, which Square acquired in 2020, according to CoinDesk. The company says it trained the model using one NVIDIA GB200 NVL72 and “hundreds of times fewer words” than large language models need. Those are company-reported training claims; the source material gives no independent measurements or comparison methodology.

Co-founder Jack Dorsey wrote that talking with Dessa feels like talking with a person, citing emotion, nuance and natural cadence. Project lead Ragavan Thurairatnam called the work “a GPT3 moment for speech” and said it could run on consumer GPUs. Neither statement substitutes for published benchmarks or outside testing. Block itself frames the research around the question of whether speech models can learn from conversational audio without converting speech into text first.

Reaction to the demonstrations has split between enthusiasm for more natural voice interaction and concern that human-like voices could lower listeners’ guard. Block has argued that voice systems should better recognize emotion, such as pain, during conversation. Critics have questioned why synthetic voices should sound human at all. Both positions concern the same capability, while the technical and safety details needed to assess Dessa remain unpublished.

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

— Block co-founder Jack Dorsey

What Independent Testing Has Not Shown

The main uncertainty is how Dessa performs beyond Block’s recordings. No outsider has publicly run the model, and there are no released benchmarks, model details or documented tests against varied speakers, accents, background noise or adversarial prompts. The source material also says Block has not published the languages supported or a public release date. It is not clear how representative the demos are of routine use.

Block has not provided details on watermarking, consent checks or safeguards against imitating a specific person. The company says Dessa chooses its own voice and personality, but that statement does not answer whether the model can reproduce a recognizable individual’s voice. Nor do the available clips establish that the system has feelings, understands speech as a person does or can be used reliably for fraud. Those questions require evidence beyond selected demonstrations.

Evidence Needed Before Wider Release

The next meaningful step would be publication of technical documentation and access for independent evaluation. A paper or model card could clarify training data, capabilities, limitations and safeguards; benchmarks could make Block’s speed and performance claims measurable. Details about consent, voice imitation and provenance would also help users and organizations judge deployment risks. Block has not announced when any of those materials or a wider release will be available.

Until then, the model should be treated as a research demonstration rather than a verified, generally available product. Organizations handling payments can review identity checks for phone requests, while individuals can avoid acting on urgent requests based only on a familiar voice. A known-number callback or an independently agreed family code word offers a check that does not depend on judging whether a voice sounds real.

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. The company’s demos show it responding with conversational cues such as laughter and interruptions.

Has Dessa been independently tested?

Not publicly, according to the source material. Block has not released a paper, model card or benchmarks, and outside researchers have not publicly run the model.

Can Dessa clone a specific person’s voice?

That is not established. Block says the model chooses its own voice and personality, but has not publicly explained whether it can imitate a specific person or what consent safeguards apply.

Does the model’s emotional voice mean it has feelings?

No. An emotional-sounding response is not evidence that a system experiences emotion or has human-like understanding. The available demos show output, not an inner emotional state.

How can people reduce the risk of voice scams?

Do not authorize payments or reveal sensitive information based only on a familiar-sounding voice. Hang up and call a number already known to you; families can also agree on a private code word for urgent requests.

Source: ThorstenMeyerAI.com