Block, the company led by Jack Dorsey, has introduced Dessa, an AI model developed to process human speech. Unlike many systems that rely on text-based training, Dessa was built from scratch using only human conversation data. The project aims to improve the natural flow of voice-to-voice interactions and enhance the AI's ability to detect emotional nuances in human speech.
The model is designed to run locally on hardware, specifically requiring a MacBook Pro with an M4 Max chip or a more powerful device. According to the development team, the primary motivations behind the project are curiosity, the pursuit of more intuitive human-AI communication, and advancements in AI safety. The team intends to address issues where current voice AI fails to recognize emotional states.
Dessa has demonstrated capabilities in various scenarios, including playing guessing games and discussing cinema. In comparative tests, the team highlighted instances where Dessa appeared more natural than OpenAI’s GPT-Live-1 voice model. The developers have announced plans to provide early access to the technology to a select group of AI safety researchers.
