Block lab tests a local model that enables synchronous voice conversation emulating human emotion and warns of deception risks

Early experiments with the Disa model, developed by Block’s research lab, reveal a tangible shift in smart-voice engineering as the new system abandons intermediate text processing in favor of direct, end-to-end training on human speech. The model announced by Jack Dorsey operates with fully synchronous send-and-receive, allowing it to listen and speak at the same moment just like humans, while displaying live expressions such as laughter, pauses, sighs and sad tones without needing massive cloud servers to run it.

The new voice model reshapes the relationship between machine and human emotions through computational training performed on a single Nvidia graphics processing unit.In today’s prevailing models, systems convert speech to written text and then feed it to a large language model to retrieve an answer before converting it back to spoken language, a processing chain that Block says wastes huge amounts of fine-grained information that the raw text does not carry, such as the speaker’s tone, tension and mental state. While systems such as ElevenLabs v4 or Fish Audio Drama 3 resort to textual tags to steer emotions, the company claims that emotions in its model are embedded within the model’s neural fabric itself, relying on directly heard dialogue data to build its knowledge.

Recordings released by the lab’s researchers, including Raghavan Thooriratnam and Aiden Polat, showed a voice robot expressing fatigue after work in a strained tone that mimics genuine exhaustion, and in another conversation it spoke of feeling mental relief in a wavering voice that nearly broke under emotion, as well as its speed in engaging interactive guessing games.The model’s operational strength lies in its ability to run locally on personal computers such as MacBooks equipped with M4 Max processors or equivalents.Company researchers described this as a pivotal moment comparable to the emergence of the GPT-3 model in the trajectory of voice systems, because it shifts live-conversation technology from supercomputing monopolies to independent physical devices.

Despite this flexibility, the company blocked public access to the model and issued an open invitation to cyber-security researchers to test its vulnerabilities and security limits. This restraint reflects a precise awareness of the risks posed by generating voices that can closely mimic human feeling, as the likelihood of using voice agents to deceive listeners or mislead them in interactive conversations increases, especially since the company primarily operates in financial services and sells solutions to small businesses, and employs voice AI tools in its Square platform to receive commercial requests and reduce operating costs.

This shift carries clear strategic implications for technology firms in the Gulf, Egypt and the broader Arab region. Enabling voice models to run locally on office computers eliminates the rising costs that call centers and digital commerce platforms pay to cloud-server providers by the minute and second, and also resolves the regulatory-compliance and data-sovereignty dilemma, as banking and service sectors can process sensitive customer voice calls within their premises without routing audio through external cloud gateways.The equation simultaneously demands a urgent review of regional banks’ voice-verification and fraud-prevention systems.Now that modest personal hardware can generate convincing human-like interaction that outpaces traditional detection tools.

We see in this development an indicator of the shift in AI processing from remote data centers to devices you hold in your hand or place on your desk. Bypassing text mediation not only shortens response time in voice interaction, but also presents businesses with a new technical reality that requires rebuilding customer-service interfaces and carefully balancing reduced field-operation costs against securing communication channels from targeted manipulation attempts.