Block’s Dessa Makes A Human-Sounding AI Voice. What Does That Show?

🔍 Read the full analysis: Block’s Dessa Makes A Human-Sounding AI Voice. What Does That Show? on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

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 listen and talk at the same time, with laughter, pauses and interruptions that resemble human conversation. The release matters because natural-sounding voice systems could make spoken AI more useful, but also make it harder to recognize impersonation; so far, the public evidence consists of Block-selected recordings, not independent tests.

Block posted a project page for Dessa on October 23. The company says the model was trained from scratch on human conversation, without text or pretrained models, and can run locally on a MacBook Pro M4 Max or better. Block says inference is faster than real time but has not provided a latency figure. It also says training used one NVIDIA GB200 NVL72 and far fewer words than large language models typically require.

The demos show Dessa laughing mid-sentence, responding to a forgotten film title and reacting to a surprise-anniversary scenario. Block describes the clips as live recordings, saying nobody steered the conversations. Co-founder Jack Dorsey wrote that talking with the system feels like talking with a person, citing its emotion, nuance, pauses and cadence. Those descriptions are Block’s claims about its own model, not results from an independent evaluation.

Outside coverage has been more qualified. The Verge described some exchanges as occasionally awkward and uncanny, while finding Dessa more natural than another model shown in the demonstrations. TechCrunch said it was unclear whether emotion might be guided by tags. As of October 26, Block had not made public a technical paper, model card, benchmark results, model weights or API. Access was limited to a waitlist and a small group of AI safety researchers, according to the source material.

Natural Voices Raise Fraud Risks

Dessa’s demos point to a potential improvement in voice interfaces: a system that can respond to pauses, interruptions and emotional cues may feel less like a rigid sequence of spoken commands. Block argues that speech systems often miss emotion in a person’s voice, which can make them insensitive to a user’s situation. Whether Dessa reliably understands those cues outside selected demos remains untested.

The same conversational realism could make voice impersonation more persuasive. The FBI’s 2025 Internet Crime Report, published in April 2026, counted 22,364 complaints mentioning AI and $893 million in reported losses. It also reported more than $5 million in losses from distress scams involving cloned voices of loved ones. Those figures describe reported complaints and losses; they do not show that Dessa was used in a scam. The relevance is that convincing speech is already part of a broader fraud risk, and more natural interaction could make familiar warning signs less reliable.

For users and organizations, a familiar-sounding voice should not stand alone as proof of identity, especially when a caller demands money or urgent action. Independent verification through a known number or a separate channel can reduce reliance on voice recognition. Any safeguards specific to Dessa, such as watermarking or controls against imitating an identifiable person, were not described in the available material.

A Research Model, Not a Release

Dessa is associated with a small research team at Block, the company behind Square and Cash App. The source material says the team was founded by former members of Dessa, a Toronto deep-learning startup acquired by Square in 2020. Block’s announcement presents the project as research rather than a broadly available product. It describes the system as speech-only and full-duplex: it can listen while speaking, rather than waiting for a spoken exchange to be converted into text and then answered.

That distinction is central to the company’s pitch. Block says speech-to-text systems can discard qualities such as sarcasm, laughter or a frustrated sigh. The demos are meant to show a model responding to those features more naturally. But recordings chosen by the developer cannot establish how well the model performs across ordinary conversations, accents, languages, difficult audio or adversarial situations. Block has identified learning new languages as an open question.

Block lead Ragavan Thurairatnam called the project “a GPT-3 moment for speech” and said it could run on consumer GPUs. That is a characterization by the project’s lead, not a benchmarked finding. The source material also notes that Block has not disclosed the model’s size or a public release date. The available demonstrations can establish what the company chose to show, not the system’s full capabilities or limits.

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

— Jack Dorsey, Block co-founder

Independent Tests Still Missing

No outside researcher has publicly tested Dessa, and Block has not released the materials needed to reproduce or benchmark its claims. The model’s size, detailed performance, failure rates and latency remain undisclosed. The demonstrations do not establish how it handles a wide range of speakers, languages, noisy calls or sensitive scenarios.

It is also unclear whether emotional qualities in the demos arise entirely from the model or are influenced by tags or other controls. Block says Dessa picks its own voice and personality, but the available material does not explain whether it can imitate a specific person, how consent is handled, or whether outputs are watermarked. No public release date or language list was provided. These gaps make it too early to judge the system’s reliability or its practical safeguards.

Waitlist and Safety Review

Block’s stated next step is limited access through a waitlist and early access for a small group of AI safety researchers. The company has not announced when wider access, model weights or an API might become available. Independent testing and technical documentation would help clarify whether the demo behavior holds up beyond the recordings and what protections accompany the model.

Until then, the public can assess only the company’s demonstrations and the limited outside reporting on them. People receiving urgent requests for money should verify them using a number they already know or another trusted channel, rather than relying on a familiar voice alone. Whether Block will publish information on consent, impersonation controls, watermarking and language support remains unknown.

Key Questions

What is Block’s Dessa?

Dessa is a speech-to-speech research model from Block. The company says it was trained on human conversation and can listen while speaking, with behaviors such as laughter and conversational interruptions.

Has Dessa been independently tested?

Not publicly, according to the material available as of October 26, 2026. Block has shown selected recordings, but has not released a technical paper, benchmarks, model weights or an API.

Does Dessa prove that AI understands emotion?

No. The demos may show emotionally expressive speech, but that does not establish that the model experiences emotion or reliably understands a speaker’s feelings.

Could human-like AI voices be used in scams?

More natural speech could make impersonation harder to detect, but there is no evidence in the source material that Dessa has been used in a scam. The FBI has separately reported losses from scams using cloned voices.

When will the public be able to use Dessa?

Block had not announced a public release date as of October 26, 2026. Access was described as limited to a waitlist and a small group of AI safety researchers.

Source: ThorstenMeyerAI.com