WAV, FLAC and MP3: what’s actually inside an audio file.

WAV, FLAC and MP3 are three answers to the same question: how do you put a sound into a file? WAV stores the raw measurements of the sound, untouched. FLAC stores exactly the same measurements, packed more cleverly — typically a little over half the size, and not one bit different when unpacked. MP3 works out which parts of the sound your ear is unlikely to notice, throws those away for good, and ends up a quarter to a tenth of the size. The first two are lossless. The third is lossy. Almost every practical decision about audio files follows from that one distinction.

First, what is being stored: sound as a list of numbers

Sound is air pressure wobbling. A microphone turns that wobble into a voltage, and a digital recorder does something very simple with it: it measures the voltage at perfectly regular intervals and writes each measurement down as a number. That is all a digital recording is — a very long list of numbers. The technique is called PCM (pulse-code modulation), and all three formats on this page start from it.

Two settings decide how faithful that list is:

• Sample rate — how many measurements per second. A CD takes 44,100. To capture a frequency you must sample at more than twice that frequency, and human hearing tops out around 20,000 Hz, so 44,100 covers everything audible with a little room to spare.
• Bit depth — how finely each measurement is graded. A CD uses 16 bits, which gives 65,536 possible levels. Each bit buys about 6 decibels, so 16 bits spans roughly 96 dB between the quietest and loudest sound it can describe — more than the gap between a quiet room and a jackhammer.

Multiply it out for stereo and you get the number everything else on this page is measured against: 44,100 samples × 16 bits × 2 channels = 1,411,200 bits every second, or about 10 MB a minute. That is what is on a CD, and it is what “CD quality” means.

(Why the odd figure of 44,100? Because early digital audio was stored on video recorders — the only machines of the day that could record data fast enough — and 44,100 is what you get from three samples per usable video line, on both the PAL and NTSC television systems. The number outlived the tape by forty years.)

WAV: the numbers, written down as-is

A WAV file is the list of samples with a short label stuck on the front. Microsoft and IBM defined it in 1991, and the design is about as plain as a file format gets: a 44-byte header that says how to read the samples, followed by the samples themselves, one after another, left channel then right, until the recording ends.

That simplicity is the whole appeal. There is nothing to decode, so every audio program, recorder and operating system of the last thirty years can open a WAV. It is the working format of recording studios for exactly that reason: editing software can jump to any sample instantly and change it without unpacking anything.

It is also the source of WAV’s three weaknesses:

• Size. Nothing is squeezed. Ten megabytes a minute, silence included — a minute of dead air costs exactly as much as a minute of orchestra.
• A 4 GB ceiling. The header stores the file size in a 32-bit field, which runs out at about 6¾ hours of CD-quality audio. Long or high-resolution recordings need an extended variant (RF64, or the broadcast industry’s BWF).
• Poor labelling, and no self-check. WAV has only a rudimentary way to carry the artist, album and track name, and players support it inconsistently. Nor does the file contain anything that would reveal a flipped bit — a damaged WAV simply plays the damage.

Strictly speaking WAV is a container and can hold other encodings, but in practice “a WAV” means uncompressed PCM. Apple’s AIFF is the same idea with the bytes in the opposite order.

FLAC: the same numbers, packed smarter

FLAC — the Free Lossless Audio Codec — was released by Josh Coalson in 2001 and is now an open internet standard (RFC 9639, published December 2024). It makes one promise: what comes out is bit-for-bit what went in. Decode a FLAC and you have the original WAV back, every sample identical. It just takes about half the room on the way.

The obvious question is how. An ordinary ZIP barely dents a WAV file, because ZIP hunts for repeated strings of bytes and audio almost never repeats exactly. FLAC exploits something different: sound is predictable. A waveform is a smooth curve, so if you know the last few samples you can make a very good guess at the next one.

So FLAC works in three steps:

1. Cut the audio into blocks — typically 4,096 samples, a little under a tenth of a second.
2. Find a formula that predicts each sample from the ones before it, tuned to that block. Store the formula.
3. Store only the misses — the difference between each prediction and the real sample. Because the guesses are good, the misses are small numbers, and small numbers can be written with far fewer bits. (The scheme is called Rice coding: short codes for small values, longer codes for the rare big ones.)

The decoder runs the same formula, adds the stored misses back on, and lands on exactly the original values. Nothing is approximated at any point. For stereo there is one more trick: instead of left and right, FLAC can store the average of the two channels and the difference between them — and since left and right are usually similar, the difference is small too.

How much it saves depends on the music. A solo piano or a quiet acoustic recording is very predictable and can shrink to 40 per cent of the WAV or less. A dense, loud, heavily compressed rock master is closer to noise and might only reach 70 per cent. Between half and two-thirds of the original is typical.

FLAC also fixes the things WAV is bad at:

• It checks itself. Every block carries a checksum, and the file header stores an MD5 fingerprint of the original, unpacked audio. Any FLAC can be tested at any time and will say plainly whether it still decodes to exactly what was encoded. (We covered how those work in how a computer knows a file isn’t broken.) For an archive, this is the feature that matters most.
• It labels itself. Artist, album, track number, cover art and anything else you care to add live inside the file, in a standard way that players agree on.
• It has no awkward size limit, and handles anything from telephone-grade mono up to 32-bit, eight-channel studio masters.

One thing that confuses people: FLAC encoders offer compression levels 0 to 8, and they are all lossless. A higher level just means the encoder spends longer hunting for a better formula, for a file a few per cent smaller. The audio is identical at every level.

The only real catch is support. Windows, Android, and most car stereos, streamers and hi-fi gear made in the last decade play FLAC natively. Macs and iPhones will play one from Finder or the Files app — but Apple’s own Music app still will not take FLAC into its library. Apple has its own equivalent, ALAC (Apple Lossless), and because both are lossless, converting between the two loses nothing.

MP3: keep what you will hear, drop the rest

MP3 starts from a completely different premise. Lossless formats ask “how do I store these numbers in less space?” MP3 asks “which of these numbers does a human being actually need?”

It came out of the Fraunhofer Institute in Germany in the late 1980s and early 1990s, was standardised in 1993 as MPEG-1 Audio Layer III, and got its file extension — .mp3 — in July 1995. Hard drives were measured in megabytes and internet connections in kilobits. A format that made a song ten times smaller changed what was possible.

It works by modelling the ear rather than the sound. The encoder works through the audio in slices a few hundredths of a second long, converting each one from a waveform into its ingredients: 576 narrow bands of frequency, from bass to treble, each with a loudness. Then it applies what is known about human hearing — psychoacoustics — to decide what can go:

• Sounds too quiet to hear at all. The ear is much less sensitive at the extremes of pitch. A faint component at 18,000 Hz is below the threshold of hearing for most adults, so it is discarded.
• Sounds hidden by louder ones. This is the big one, and it is called masking. A loud sound makes quieter sounds near it in pitch inaudible — not harder to hear, but genuinely not perceived. A cymbal crash hides the detail underneath it. A loud sound also masks whatever follows it for a moment afterwards — and, oddly, what came a few milliseconds before.

Having worked out where the ear is not listening, the encoder stores those regions coarsely or not at all, and spends its bits where they will be heard. The rounding introduces noise — but it is placed precisely where a louder sound will cover it. Finally the result is packed with a conventional lossless compressor.

How aggressive all this is depends on the bitrate, which is a budget: the number of bits the encoder may spend per second of sound.

When an MP3 does fall short, it fails in recognisable ways. Cymbals and applause turn “swishy” or watery. A sharp attack — castanets, a snare hit — gains a faint smear of noise just before it, called pre-echo, because the rounding error is spread across the whole slice of time being processed, including the quiet instant ahead of the hit. Once you know the sound you cannot unhear it, which is either a useful skill or a curse.

The honest summary: a well-made MP3 at a sensible bitrate is very, very good. The engineers tuned it for years against real listeners (famously using Suzanne Vega’s unaccompanied Tom’s Diner, because a lone voice gives a codec nowhere to hide). The patents have all expired, and it plays on essentially anything that can play an audio file. Its newer relatives, AAC and Opus, do the same job in fewer bits — but none of them is as universal.

What has changed is the reason to use it. MP3 was a brilliant answer to a storage problem, and the storage problem has gone away. That leaves it one job, which we will come to: feeding a player that accepts nothing else.

The one-way door

Here is the part that matters when you are deciding how to keep something.

Lossy compression cannot be undone. The discarded sound is not hidden somewhere in the MP3 waiting to be restored; it is gone. Convert an MP3 to FLAC or WAV and you get a file five or ten times larger containing precisely the same audio, holes and all. There are plenty of “lossless” files in circulation that are exactly this, and they are easy to spot: look at the frequency spectrum and there is a clean shelf where everything above 16,000 or 19,000 Hz has been sliced off.

Lossy on top of lossy compounds. Re-encoding an MP3 — to a different bitrate, or to another lossy format — runs the discard step again on audio that has already been through it once. The second encoder cannot tell the first one’s rounding noise from music, so it wastes bits preserving it and then adds its own. It is the photocopy of a photocopy.

Lossless formats have neither problem. WAV to FLAC to WAV to ALAC and back, a thousand times over, and the samples never change. So the rule is simple: keep a lossless master, and make any lossy copy from it. You can always turn a FLAC into an MP3. You can never turn an MP3 back into the FLAC.

Side by side

The same four-minute song, taken from a CD, saved four ways:

• WAV lossless42.3 MB
• FLAC losslessabout 25 MB
• MP3 320 lossy9.6 MB
• MP3 128 lossy3.8 MB

Which one to use

• Anything you are keeping — a CD collection, a digitised tape, a recording you cannot make again: FLAC. It is everything the original was, in half the space, labelled, and able to prove years later that it has not rotted.
• Anything you are still working on — recording, editing, mixing: WAV. Every tool reads it, nothing needs decoding, and studios usually step up to 24-bit while they work for the extra headroom. Export a FLAC when you are finished.
• MP3: only when the player gives you no choice. An older car stereo, a basic portable player, a system or upload form that accepts nothing else. Make the MP3 from the FLAC, at 320 kbps so as little as possible is lost, and treat it as a disposable copy — the FLAC stays in the archive. If the player can handle FLAC (or ALAC, on Apple gear), skip the MP3 altogether.
• Never the other direction. Do not convert MP3s to FLAC to “improve” them, and do not re-encode MP3s to a different bitrate if you can go back to the source instead.

Storage is why the advice has changed. When MP3 was designed, a single CD’s worth of WAV would not fit on a typical hard drive, so throwing most of the sound away was the only way to carry music at all. Today a thousand albums in FLAC fit comfortably on a drive that costs less than a dozen of the CDs did, and a phone has room for hundreds. Choosing MP3 to save space saves space nobody is short of — and an MP3 as your only copy gives up quality you can never get back.

Common questions