The standard way of digitizing speech: taking volume readings thousands of times per second. It's like tasting broth with a teaspoon: each sip is watery, with almost no detail. The data stream swells, though the substance is minuscule.
Scientists proposed first "simmering down" the audio signal. They pass it through filters that split it into frequency bands — resulting in a frequency map. Then they mix the bands into a few channels, distilling the essence and boosting information density. Thus, from a thin broth we get a concentrated sauce: a single drop is enough to taste. Interestingly, our ear does the same: the membrane in the cochlea splits sound into 3500 frequency bands before the brain reassembles them into speech. The method relies on the insight of Fourier: any signal is composed of simple waves. Neural networks themselves find the most meaningful combinations, filtering out noise. The result — at the same transmission rate, listeners unanimously preferred the "smartly compressed" speech over the regular. The technology will be useful for space communications and hearing aids, where bitrate and energy are worth their weight in gold.
The inner ear is a living spectrum analyzer: its membrane splits sound into 3500 bands, and only then does the brain glue them into words.