The Digital Press

This research introduces an innovative machine intelligence system that processes and generates speech directly from audio, bypassing written text entirely. It aims to serve over 700 million audio-literate people, especially in underserved rural and remote areas, by supporting languages without written forms.

Why it matters: Enables language technology access for communities without written language or digital text resources.

The big picture: Shifts machine intelligence from text-based to audio-native systems, promoting inclusivity in language tech.

Stunning stat: Over 700 million people are excluded by text-dependent AI but can benefit from audio-to-audio models.

Quick takeaway: The Multiscale Audio-Semantic Transform and fractional diffusion enable high-fidelity, textless speech generation from raw audio.