New AI Shrinks Calls and Music Into Tiny Files Without Losing Quality

A team of researchers has unveiled UniStream, an AI system that compresses audio while it is still being recorded — the technical trick behind live video calls, voice notes, and streaming music. Most audio compressors work after the fact, or they lose quality when forced to work in real time. UniStream manages both: it delivers full CD-quality sound (48 kHz) in real time, and it handles speech, music, and everyday background noise rather than just one of them.

The core idea is surprisingly human. Ordinary audio AI keeps one big dictionary of sound patterns and looks up the closest match. UniStream keeps four specialist dictionaries — think of a general practitioner referring you to one of four experts — and picks the right one on the fly. Crucially, the decoder on the receiving end can figure out which expert was used without being told, so nothing extra needs to travel over the network. A second training trick sharpens quality but is removed before use, so it costs nothing at runtime.

The results are the interesting part. At a stingy 12 kbps — roughly 5 megabytes per hour, less than a tenth of a typical MP3 — UniStream sounds about as good as EnCodec, a well-known AI compressor. At 22.5 kbps it comes within 0.03 points of Opus, the decades-old standard used by WhatsApp, Zoom, and most of the internet's phone calls, judged on speech quality. It also beat every system tested at 12 kbps or below on environmental sounds like traffic or room noise, and it runs fast enough on a graphics chip to keep up with live conversation.

The catch: this is a research paper, not a product you can download. Real-time operation needs a graphics processor, which is fine for servers but not yet for a cheap phone or earbud. And "nearly matching" Opus is not the same as beating it — the incumbent is free, already everywhere, and famously hard to budge. Still, the direction is clear: better sound from less data.