Why Audio Compresses to a Tenth: Psychoacoustics
A minute of CD-quality WAV runs 10MB; as MP3 it is 1MB, and ears barely tell. The tenfold gap is not "nine-tenths discarded" but "discarded what hearing never processed anyway".
Ears are not microphones: the blind spots of hearing
Microphones record every vibration in band; human hearing is far more selective: above 20kHz is inaudible; in quiet surroundings anything under the hearing threshold (about 30dB for much of the band) vanishes; sensitivity varies enormously across frequencies — extreme at 1-5kHz (speech), dull at the extremes.
The encoder’s first move is allocating precision along the threshold curve: inaudible bands get heavily reduced precision or zero. That information is genuinely lost — but lost without a perceptual trace. That is saving number one.
Masking: loud sounds eat quiet ones
The second saving is masking: a loud sound hides nearby weak sounds — in the instant a drum crack lands, the faint hiss beside it simply cannot be heard; frequency-adjacent sounds shadow each other, the louder swallowing the softer.
Encoders exploit this with on-demand allocation: in every time slice, the spectrum is analyzed and bits near loud components are stripped from weak ones — inaudible anyway, by definition. Music spends most of its time with a strong signal "providing cover", so masking operates almost continuously. Quiet passages (nothing but weak signal) are where bits are truly needed — which is why encoded gaps get the lowest allocation.
Bitrate tiers: how much you allow to throw
MP3/AAC bitrate tiers are budgets for what the encoder may discard: at 128kbps the budget is tight, threshold and masking decisions turn aggressive, and complex passages occasionally reveal the watery "swish" artifacts; from 192kbps the budget relaxes, decisions grow conservative, and artifacts effectively vanish; 320kbps is transparent for nearly everyone and everything.
Reframing compression decisions with the theory: speech is narrow-band and heavily masked — 64kbps mono is clean; classical music’s wide dynamics and quiet passages demand 256kbps and up. The spectrogram makes the theory visible — that clean cutoff above 16kHz is the encoder discarding "highs you cannot hear", filed as evidence.