Release of New AI Voice Model
Block Inc., the fintech company led by Jack Dorsey, has released a research preview of a new artificial intelligence voice model named Dessa. The model is currently hosted on the company’s official website (block.xyz/dessa) and is being positioned as a significant development in the firm’s generative AI research.
According to technical details released alongside the announcement, Dessa is a full-duplex voice conversation model trained exclusively from human speech. Unlike many contemporary large language models, the model was trained from scratch using audio data without the use of text or pre-trained models. This architecture allows the model to both listen and speak simultaneously.
Research Preview and Technical Specifications
The model is currently in a research preview phase. Block has indicated that the model is capable of running locally on high-end consumer hardware, specifically mentioning compatibility with MacBook Pro M4 Max devices or higher, where it can operate at speeds faster than real-time. The training process for the model reportedly utilized a single NVIDIA GB200 NVL72 system.
Project lead Ragavan Thurairatnam has described the development as a “GPT-3 moment” for the field of voice technology. At present, the model has not been opened for general public use; instead, Block has released demonstration recordings and provided early access to a limited number of AI safety researchers.
Market Context
The release of Dessa comes amid a broader industry shift toward “agentic” AI and specialized voice interfaces. While Block focuses on this speech-centric approach, other major players are pursuing different trajectories, such as Anthropic’s Claude Duet, which focuses on professional productivity, and OpenAI’s Dots, which is designed for multi-step, long-running projects.
As the model remains in a research capacity, the full scope of its commercial application within the Block ecosystem remains unconfirmed. The company has not yet released formal technical benchmarks or performance metrics regarding the model’s latency and accuracy.