How Can FLAC Compression Be Lossless? A Counterintuitive Explanation

The counterintuitive question

The sentence "the file shrinks by half but nothing is lost" sounds illogical the first time you hear it. If you halve something, surely something must have gone missing?

Actually, no. And to understand why, you first need to see how wasteful uncompressed data is.

Why is WAV so wasteful?

A WAV file contains the instantaneous amplitude values of a waveform, measured 44,100 times per second. Each measurement is written as a 16-bit number — a value between -32768 and +32767.

Now the critical point: neighboring measurements are not very different from one another.

A sound wave is a physical phenomenon; air pressure does not leap from -30000 to +30000 in an instant. Consecutive samples are generally close to each other. But WAV does not care — it writes every sample at a full 16 bits, as though it were a completely independent number.

That is where the compressible space lies. The data occupies far more room than the information it carries.

FLAC's core idea: record the difference

Let's start with a simple example. Suppose you have these consecutive samples:

1000, 1005, 1009, 1012, 1014, 1015

If you write them as they are, you have to spend a full 16-bit slot on each. But if you write the differences between them:

1000, +5, +4, +3, +2, +1

Apart from the first value, they are all single digits. Writing these numbers does not require 16 bits — 4 or 5 bits is more than enough. And no information was lost: if you know the first value and the differences, you can reconstruct the original sequence exactly.

What FLAC does is a far more sophisticated version of this.

Linear prediction: predict, then write the difference

FLAC does not simply take the neighboring difference. Instead, it builds a prediction model for each block of audio.

The model answers this question: "looking at the preceding few samples, how do I best predict what the next sample will be?" This is called linear prediction. The encoder computes the coefficients that work best for that block.

Then it records:

1. The model coefficients (a few numbers per block — very little space).
2. The residual values: the difference between the model's prediction and the actual sample.

If the model works well, the residuals are small numbers. And small numbers need few bits.

Why is there no loss?

The answer to that question is that the operation is performed with exact arithmetic.

The decoder reads the model coefficients stored in the file and makes exactly the same prediction the encoder made. Then it adds the recorded residual on top of that prediction:

actual sample = prediction + residual

No rounding, no approximation, no "close enough." Because integer arithmetic is used, the result is the original sample itself. Every time, for every sample.

This is precisely the difference from lossy formats. MP3 says "if I discard this information, nobody will notice" — and genuinely discards it. FLAC says "I can write this information more concisely" — and discards nothing.

Coding residuals efficiently

In the final step, FLAC writes the residual values using an entropy coding. This coding is optimized for a distribution where small values are frequent and large values are rare — which is exactly how audio residuals are distributed.

So frequently occurring small values are written with very few bits, and rarely occurring large values with more. The average bit count drops substantially.

What does the compression level do?

FLAC's 0-8 compression level determines how much effort the encoder spends:

• How many different prediction models will it try?
• How will it select block sizes?
• Will it exploit the relationship between channels (in stereo recordings, mid-side coding usually compresses better than left-right)?

At a high level the encoder tries more options and picks the best. This takes longer and produces a slightly smaller file.

But quality never changes. Files produced at level 0 and level 8 yield bit-for-bit identical PCM data when decoded. The level is not a quality setting; it is a computation budget setting.

On the decoding side, the level difference is essentially imperceptible; a file at any level decodes quickly.

Why does the compression ratio vary?

You now know the answer: predictability.

• Speech, acoustic solos, quiet passages: Smooth waveform structure, accurate prediction, small residuals. Compresses very well.
• Classical music, room recordings: Wide dynamics but regular structure. Compresses well.
• Dense rock/electronic mixes, heavy mastering: The signal is saturated and complex; prediction struggles. The gain drops.
• White noise: Unpredictable by definition. Barely compresses at all.

If your recording compressed poorly, it does not mean you picked the wrong setting — that is simply its nature.

The fundamental distinction from lossy compression

It is worth laying out what we have covered in comparative terms, because this is usually where the confusion originates.

A lossy encoder like MP3 asks: "if I discard this information, will the listener notice?" Using models of human hearing, it identifies components unlikely to be heard and genuinely deletes them. The gain is large — the file shrinks several times over — but the deleted data remains nowhere in the file. Even if you later convert that MP3 to FLAC, the information does not return, because there is nothing to return.

A lossless encoder like FLAC asks a different question: "how can I write this information more concisely?" It discards nothing, only makes the expression more efficient. The gain is more modest — roughly half the size — but the round trip is complete.

Both approaches are legitimate; they just serve different purposes. A lossy format makes sense for portable listening; a lossless format is required for archiving and production. The point not to confuse: moving to a lossless format does not recover data that was previously lost in a lossy format. It only protects everything from that point forward.

The difference between FLAC and ZIP

If you ZIP a WAV file, you have also performed lossless compression. But the gain is usually very small.

The reason: ZIP is general purpose and looks for repeating byte sequences. It works wonderfully on a text file where the word "the" appears hundreds of times. In audio data, byte sequences that repeat exactly are almost nonexistent.

FLAC, by contrast, knows what the data is. It looks not for byte patterns but for waveform structure. That domain knowledge explains the large performance gap between them.

A bonus: integrity verification

FLAC writes checksums for each block and for the entire file. This is a genuine advantage that WAV lacks.

When you open an archive drive years later, you can definitively verify whether your FLAC files are still intact. In WAV, a corrupted region silently produces wrong audio; to notice it you would have to listen through the file from beginning to end.

Summary

FLAC saves space without discarding anything from the audio data because uncompressed PCM is already wasteful: it writes every sample at full size as though it were independent, when in fact neighboring samples are tightly related. FLAC captures that relationship with a prediction model and records only the prediction error. When decoding repeats the same prediction and adds the error back, the original data returns exactly. That is why you can use our WAV to FLAC converter with complete confidence: the only thing you lose is disk space.