Block’s new voice model sounds eerily human
AI-summarised brief · reviewed before publication
Block, the company founded by Jack Dorsey, unveiled Dessa, a full‑duplex speech‑to‑speech model that runs on a single NVIDIA GPU. Unlike typical voice AIs that convert speech to text before generating audio, Dessa is trained entirely on human speech and produces laughs, pauses, and even cries without external emotion tags. The model can listen and speak simultaneously, mirroring OpenAI’s GPT‑Live, and is claimed to sound “eerily human” in demos that show varied cadences and emotional states. Block says the system can operate locally on devices such as a MacBook M4 Max, reducing reliance on cloud services. The company has opened a safety‑research program but has not released public access, positioning Dessa as an on‑device alternative for finance and small‑business applications.
💡 Why It Matters
- · By delivering human‑like vocal nuance without cloud dependence, Dessa could redefine how businesses embed conversational AI into everyday devices, raising both user experience and privacy stakes.
