Some Sounds Disappear Without Being Noticed有些聲音，消失了也不會被發現

At the time, I treated MP3 as a problem of signal processing and engineering trade-offs: how much numerical precision could be surrendered before the loss became audible? Where would errors accumulate once the representation changed? Years later, what the project left behind was more than an implementation technique.

It posed a more interesting question: when resources are limited, what should we keep?

1,411,200 bits

One second of uncompressed stereo CD audio.

MP3 does not try to preserve every one of them. It models how we hear first, then leaves less room for what is hardest to notice. The sound still sounds like itself — with far less data behind it.

Compression is not about making everything smaller.
It is a decision about what need not remain.

None of this is arbitrary deletion. It rests on psychoacoustics, mathematical models, signal transforms, and a great deal of listening. MP3’s achievement is not forgetting a lot — it is forgetting precisely.

It does not hear less. It knows what you cannot hear.

The shorthand explanation says MP3 removes frequencies humans cannot hear. That is not exactly wrong, but the real mechanism is finer: whether a sound is audible also depends on when it occurs and what is playing beside it. For each short frame of sound, the encoder takes four steps.

01 · Split — a moment into frequencies. The encoder decomposes the frame into frequency components and watches how energy is distributed across time and frequency.

02 · Estimate — draw the hearing threshold. A psychoacoustic model estimates which components are easy to hear and which will be masked by stronger neighbours.

03 · Trade — spend bits where they matter. Perceptually sensitive components receive more precision; masked components are represented with fewer bits, and quantisation noise is pushed, as far as possible, below the masking threshold.

04 · Pack — record the common more briefly. The quantised values then pass through entropy coding: frequent values get shorter codes, rare values longer ones. This step loses nothing further — but nothing discarded earlier can ever be played back.

One voice once defeated MP3.

During development, American singer-songwriter Suzanne Vega’s near-a-cappella recording of “Tom’s Diner” became a famous test track. With little accompaniment to hide behind, its exposed vocal texture readily revealed coding artefacts; MP3’s core researcher Karlheinz Brandenburg listened to it over and over while tuning the system, and the song was later called the “mother of MP3”.

The story is a reminder that mathematical models must return to human ears. An encoder can calculate a masking threshold; someone still has to hear when the model is wrong.

Low-bitrate MP3 sometimes turns metallic, warbles, or produces pre-echo — a smear of sound leaking out just before a handclap or drum hit has happened. These artefacts say something honest: lossy compression is not the objective removal of the unimportant, but a bet, placed by a model, on which losses will go unnoticed.

What has MP3 forgotten?

Does it simply delete everything above 20 kHz? No. Low-pass filtering may be part of some encoder or bitrate strategies, but MP3’s core is not a blunt cut. It estimates frequency and temporal masking within short frames, then allocates its limited bits.

Can the discarded data be recovered? Not exactly. MP3 is lossy by design; the decoder reconstructs an approximation. Bit-perfect recovery of the original PCM needs a lossless format such as FLAC.

Is 128 kbps a fixed level of quality? Not necessarily. Material, encoder, settings and listening environment all shape the result. A bitrate is a similar data budget, not a guarantee of identical perceived quality.

Why did MP3 succeed, rather than merely work? Because it solved the right constraints at the right time: tiny hard disks, 28.8 kbps modems, and then portable players. Compression turned “sending a song” from impractical into a wait you could accept.

Memory keeps content. Forgetting makes room for structure.

MP3 does not know what a song means to you. It only knows that at a given instant, some signals are likely to be covered by others.

But it leaves an observation worth carrying beyond audio: receiving more is not the same as understanding more. When attention is limited, selection is not a flaw in the system. Selection is the system.

Compression is a judgement about what deserves to remain,
written down as executable technique.

So missing a few days of information is not necessarily falling behind. Sometimes it is simply reallocating limited attention to what has not yet been absorbed.

Sources & notes

1. Fraunhofer IIS — The mp3 History: Timeline
2. Fraunhofer IIS — Development
3. Fraunhofer IIS — Technology
4. MPEG — MPEG-1 Audio
5. Fraunhofer IIS — 30 Jahre .mp3
6. Fraunhofer IIS — Dieter Seitzer: pioneer and founding director
7. Apple Newsroom — Apple Presents iPod (2001)

File sizes are theoretical estimates, excluding container/metadata overhead. The listening demonstration synthesises tones in the browser to convey the intuition of masking; it is not the full psychoacoustic model of an MP3 encoder.