Block unveils Dessa, a speech-only AI model trained on human conversation

Block has announced the release of Dessa, an artificial intelligence model designed to process and generate speech directly as audio. Unlike traditional systems that convert speech to text before responding, Dessa operates as a true speech-to-speech model. The company states that the model was trained entirely on human conversation without the use of text data. This approach allows the system to handle full-duplex communication, meaning it can listen and talk simultaneously.

The model is capable of running locally on a MacBook Pro equipped with an M4 Max chip or better. Block revealed that Dessa was trained from scratch using a single NVIDIA GB200 NVL72 computing system. The company claims that emotion is built into the model’s architecture rather than being steered by external voice tags. Jack Dorsey noted on X that interacting with the model feels like talking with a person who displays emotion, nuance, and natural cadence. However, Block shared examples of the model making mistakes, such as initially mistaking a speaker for someone named Ronaldo.

Reactions to the announcement have been mixed. Some observers expressed amazement at how human the voice sounds, while others raised concerns that the technology could fuel scams and deception. The team stated they built Dessa out of curiosity, a desire for better human-AI communication, and safety concerns. Block has not disclosed the model size, benchmarks, or a specific release date for the product.

Dessa remains a research preview and is not yet publicly available. Block plans to provide early access to a small group of AI safety researchers. The company is currently looking for AI security researchers to test the system before any broader rollout.