Block Unveils Humanlike Dessa Voice AI as Scam-Call Concerns Grow

A voice system that laughs in mid-sentence, reacts to a surprise and keeps responding through interruptions can make a conversation feel natural. It can also remove familiar cues people use to decide whether a caller is real. Block has unveiled Dessa, a voice artificial-intelligence model whose humanlike conversational behavior raises a practical concern: synthetic scam calls may become harder to recognize by sound alone.

Block says its demonstrations were recorded live without a person directing the model’s responses. Dessa laughs during one exchange and responds to a surprise with apparent emotion in another. That behavior should not be confused with evidence of genuine feeling, nor has Block provided independent listening tests showing that people mistake Dessa for a human. The demonstrations nevertheless expose the underlying human-factors problem: laughter, pauses, cadence and quick reactions are among the signals listeners commonly associate with a person on the other end of a call.

Direct audio changes the conversation

Many voice assistants operate as a chain of separate functions. Speech is transcribed into text, a language model produces a text response, and another system converts that answer back into speech. Each handoff can introduce delay or strip away timing and vocal detail.

Block says Dessa instead processes audio directly and was trained entirely on human speech rather than text. It can listen while speaking, a capability often described technically as full-duplex operation. In practical terms, the model can respond to an interruption, laughter or a change in tone without waiting rigidly for one speaker’s turn to end.

Those characteristics help explain why Dessa can sound more conversational than a conventional voice interface. They also make verification harder. A hesitant pause, an overlapping reply or a well-timed laugh can no longer be treated as reliable evidence that a caller is human. For anyone helping a parent or grandparent screen persistent scam calls, the important question is no longer whether a voice sounds robotic. It is whether identity can be confirmed through a separate, trusted channel.

Local operation lowers an important barrier

Block says it trained Dessa from scratch on one Nvidia GB200 NVL72 computing system. That is substantial training infrastructure, but running the finished model is a different engineering task. According to the company, Dessa is compact enough to operate locally on a MacBook Pro equipped with an M4 Max chip or better.

Local operation matters because a voice model would not necessarily need to send every exchange to a remote data center. That can reduce latency and dependence on network connectivity while keeping processing on the device. It also broadens the potential deployment path beyond organizations operating large cloud systems, although an M4 Max MacBook Pro remains high-end consumer hardware rather than an entry-level machine.

Several important boundaries remain unknown. Block has not specified whether Dessa is publicly available, what access controls or safeguards it may use, or whether generated calls would carry an audible or machine-readable disclosure. The company also has not released independent results measuring whether listeners can distinguish Dessa from human speech. Calling it a public product or a research preview would therefore go beyond its confirmed status.

Detection is becoming an adaptive contest

Automated synthetic-speech detection offers one possible layer of defense, but current research shows why a single detector may not be enough. A recent IEEE study of synthetic-speech detection combined multiple detection models and changed their weighting for each utterance. Tested on the ASVspoof 5 benchmark and several out-of-domain datasets, the dynamic approach outperformed conventional ensemble methods and adjusted which detector it trusted according to the input.

That result is relevant to systems such as Dessa because synthetic voices, codecs, compression and real-world call conditions vary. A detector trained on one generation method or dataset may perform poorly when the domain changes. Detection therefore has to generalize to unfamiliar speech generators rather than merely recognize artifacts from systems it has already encountered.

Technical detection is also only one part of a safer communications system. Disclosure rules, service-provider screening, account controls and identity checks can reduce reliance on a listener’s intuition. Human procedures remain important as well: a convincing voice should not, by itself, authenticate an urgent request involving money or sensitive information.

Dessa’s most consequential feature may not be laughter or apparent surprise in isolation. It is the combination of direct-audio conversation, interruption handling and local operation on high-end consumer hardware. Block has shown how natural synthetic speech is becoming; it has not yet shown how access, disclosure and authentication will keep pace.

By Thomas Caldwell — AMI’s senior editor for mechanical and mobility engineering, covering vehicle electronics, systems integration, electrification, chassis systems, propulsion, and safety policy.