Why AI Voices Can Sound Personable Without Proving Anything

🔍 Read the full analysis: Why AI Voices Can Sound Personable Without Proving Anything on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get smart everyday buys 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 released demonstration recordings of Dessa, a speech-to-speech research model that the company says can listen and talk at the same time, respond to emotion and run locally on powerful Mac computers. The clips sound notably conversational, but no independent researcher or reporter has tested the model directly, and Block has not published a technical paper, benchmark results, or public access. That gap matters as realistic voices are already used in fraud attempts.

Block published a project page for Dessa on October 23. The company says it trained the model from scratch using human conversation, without text or pretrained models. It describes Dessa as full-duplex: the system can listen while speaking, interrupt, and provide conversational responses such as laughter or backchanneling. Block also says it can recognize and produce vocal cues such as sarcasm and sadness.

The recordings on Block’s page include a model laughing mid-sentence, helping a user who cannot recall a film title, and responding to news of an anniversary surprise. Block says the clips are live conversations and that nobody steered the model. These remain company-selected demonstrations, not results from a controlled outside evaluation. The company has not released model weights or an API, and access is limited to a waitlist and early access for a small group of AI safety researchers.

Block says inference runs faster than real time on a MacBook Pro M4 Max or better, but has not supplied a latency figure. It also says Dessa was trained using one NVIDIA GB200 NVL72. The model’s size, benchmark performance, language coverage and release date have not been disclosed, according to reporting cited by the source material. Those missing details make it hard to compare the claims with other speech systems or assess how the model behaves beyond the published clips.

Why Human-Sounding Calls Raise Risk

A voice that handles pauses, interruptions and emotion can make an automated interaction easier and more natural. The same qualities could also make an impersonation harder for a listener to identify. Conversational realism is not proof of identity: a familiar-sounding voice on a call should not, by itself, authorize a payment or reveal sensitive information.

The source material cites the FBI’s 2025 Internet Crime Report, published in April 2026, which counted 22,364 complaints mentioning AI and $893 million in reported losses. The FBI also counted more than $5 million in losses from distress scams using cloned voices of loved ones. These figures describe complaints and reported losses, not a measurement of Dessa’s capabilities or evidence that it has been used in a scam.

For businesses, the practical concern is that voice-based checks may not be dependable when synthetic speech can sound more natural. The source recommends confirming payment requests through a known, separate channel and setting clear rules for synthetic voices. For households, a pre-agreed family code word and calling back using a saved number can help check urgent requests. These steps address impersonation risks that exist independently of Block’s model.

What Block Has Released So Far

Block, the company behind Square and Cash App, introduced Dessa through a project page and public recordings rather than a downloadable model or technical report. Co-founder Jack Dorsey described the interaction as person-like, while project lead Ragavan Thurairatnam called the work “a GPT3 moment for speech.” Those are the speakers’ characterizations, not independent findings.

Outside coverage has been more cautious. The Verge described some moments as “a little awkward and uncanny,” while also finding Dessa more natural than another model featured in a comparison. TechCrunch said it was unclear whether emotion in the recordings might be guided by tags. CoinDesk reported that Block had not disclosed the model’s size, benchmark results or a public release date. These observations concern the available clips; they do not amount to a direct test of the underlying model.

The source also points to regulatory and organizational questions. It says the EU AI Act’s Article 50 rules apply from August 2, 2026, and recommends disclosure when people interact with AI-generated voices. The material does not establish how those rules apply to every use of Dessa. Block’s published claims, the limited recordings and the absence of outside access leave that assessment open.

“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 Are Still Missing

As of October 26, Block had not published a technical paper, model card, license, benchmark results or public release date. No outside party had run the model, so its performance on unfamiliar prompts, interruptions, noisy calls or languages beyond any demonstrated examples remains unverified. The company describes training and local operation, but has not provided enough technical detail to independently assess those claims.

It is also unclear whether Dessa can imitate a particular person, what consent checks or safeguards it uses, or whether its output is watermarked. Block says the model selects its own voice and personality, but that does not answer how voice identity is protected in practice. The available clips do not establish that the system has been used in fraud, nor do they show how often it makes errors or how it handles high-stakes conversations.

Access and Safeguards to Watch

The next useful evidence would be independent testing, a technical report and details about access, voice imitation controls, consent and disclosure. Block has not announced when a wider release might happen, so a date for those developments is unknown. Until researchers can examine the system directly, claims about its speed, conversational ability and safety should be treated as company-reported, not independently confirmed.

For readers and organizations, the immediate response does not depend on whether Dessa reaches the public. Verify urgent requests through a contact method already known to be genuine, rather than calling back a number supplied during the same conversation. Businesses can require written approval or a second channel for transfers, and families can agree on a code word. These measures do not prove who is speaking; they provide a separate check when a voice alone is no longer persuasive evidence.

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 says it can listen and speak at once and respond to vocal emotion.

Has anyone independently tested the model?

Not directly, according to the source material as of October 26, 2026. Reporters and listeners have assessed Block’s recordings, but public model access and independent benchmark results are not available.

Does the demo prove that Dessa can clone a specific person?

No. The recordings do not establish whether Dessa can imitate a particular person. Block says it selects its own voice and personality, but details about voice imitation and safeguards remain undisclosed.

How can people check an urgent voice request?

Hang up and call the person using a number already saved or otherwise independently verified. Families can also use a pre-agreed code word; organizations can require a second channel for payment approvals.

Source: ThorstenMeyerAI.com