Jack Dorsey Unveils Block’s Speech Model for Two-Way Audio

Block co-founder Jack Dorsey announced Dessa on Oct. 23, introducing a speech model designed to take in audio and generate spoken replies. Block describes the system as full-duplex, meaning it can listen and speak at the same time. The model centers on a two-way audio exchange.

Block says it trained Dessa from scratch on human conversation, without text or pretrained models, using one NVIDIA GB200 NVL72 system. The project page does not list the model’s size, training-data volume or benchmark results.

Block presents Dessa partly as a safety research project, arguing that voice can carry signs of distress or pain that may be lost when speech is converted to text. Dorsey described the demonstrations as “full of emotion, nuance, pauses, and a natural cadence.” Block is offering early access to a small group of AI safety researchers, and its project page provides no public release date.

A separate 2024 study by University of Illinois researchers tested voice-enabled GPT-4o agents equipped with browser tools in simulated scams. The agents achieved a 36% overall success rate. In one crypto-transfer scenario, they had a 40% success rate across five runs. The study did not test whether the agents could persuade victims.

Those results concern the GPT-4o system used in the experiment, not Dessa. The study provides a separate example of voice-enabled AI used in simulated scams; it does not measure Block’s model.

Block’s portfolio includes Square, Cash App and Bitkey, a self-custody wallet for Bitcoin (BTC). The Dessa announcement did not describe an integration with those products.