Block, the financial technology company co-founded by Jack Dorsey, has introduced Dessa, a speech-to-speech artificial intelligence model designed for real-time two-way conversation. The system processes audio input and generates spoken responses without converting speech to text at any stage.

According to the company, Dessa was trained entirely from scratch on recordings of human conversation. No text data or preexisting language models were used in the process. Training took place on a single NVIDIA GB200 NVL72 system, a liquid-cooled rack containing 72 Blackwell GPUs.

Full-Duplex Design and Local Performance

Dessa operates in full-duplex mode, allowing it to receive audio while simultaneously producing spoken output. This enables natural conversational behaviors such as interrupting, completing sentences, and using short backchannel responses like acknowledgments during ongoing speech.

The model runs inference faster than real time and can operate locally on a MacBook Pro equipped with an M4 Max chip or a higher-specification device. Block states that the system captures and reproduces emotional cues, including laughter, sarcasm, sighs, and shifts in tone that are typically lost when speech is transcribed into text.

Dorsey publicly announced the project, noting the demonstrations display natural conversational qualities. Project lead Ragavan Thurairatnam described the release as comparable to a GPT-3 moment for speech AI.

Research Focus and Limited Access

The model is currently offered as a research preview. Early access has been granted to a small group of AI safety researchers, with a public waitlist available for others. No model size, training dataset volume, formal benchmarks, API, or general release date have been disclosed.

Block’s AI research lab, staffed by a small team including former members of the Toronto deep learning startup Dessa acquired by Square in 2020, reports directly to Dorsey. The lab positions the work as addressing limitations in existing voice systems that flatten audio into text and may miss emotional signals such as distress.

Demonstration recordings released by the company show interactions in which the model finishes incomplete thoughts, responds with appropriate emotional inflection, and maintains conversational timing. Subtitles were added after the recordings; the sessions themselves were not steered by operators.

Company Context

Block’s broader operations include Square, Cash App, and the Bitcoin self-custody wallet Bitkey. The Dessa announcement did not outline any planned integration with those products. The project remains focused on research exploration of speech-only training methods and the behaviors that emerge from models developed without text foundations.

No independent evaluations of performance, latency figures, or supported languages have been published. The company has indicated that additional recordings will be shared in the coming period as the research preview continues.