There is a question that sounds like it should have a simple answer and does not: how much of the music is still in the file?
For a lossless format the answer is all of it, exactly. For a lossy one the honest answer is “enough, probably, for how you are going to listen to it”, and the gap between those two statements is where every argument about audio quality actually lives.
Two different jobs, not two quality levels
Lossless compression makes a file smaller without changing what comes out. FLAC works the way ZIP works: it finds structure and redundancy in the data and encodes it more efficiently. Decompress it and you get back the identical sample values you started with, bit for bit. A FLAC of a CD track is typically around half the size of the raw audio, and it is the same recording.
Lossy compression makes a file dramatically smaller by deciding what you will not miss. MP3, AAC and Opus are built on psychoacoustic models: descriptions of what human hearing actually resolves. The encoder analyses the signal, identifies parts the ear is unlikely to notice, and simply does not store them. Decompress it and you get something close to the original, never the original.
The size difference is the reason anyone tolerates the second approach. Uncompressed CD-quality stereo runs at 1,411 kbps, which is about 10 MB per minute; a five-minute song is 50 MB. The same song at 320 kbps is around 2.3 MB per minute, roughly 23% of the original data. At 128 kbps it is about 9%.
That is not a small saving. It is the difference between a library that fits on a phone and one that does not, and it is why lossy formats won.
What the encoder actually removes
It is worth being concrete, because “removes what you cannot hear” sounds like marketing and is in fact a description of two measurable phenomena.
Frequencies outside your hearing. Human hearing tops out somewhere between 15 and 20 kHz, declining with age. Content above that range is real, present in the recording, and inaudible to you, so it goes first. This costs nothing perceptually and saves a surprising amount, because high frequencies are expensive to encode.
Sounds masked by louder sounds. This is the clever part. When two sounds occur close together in frequency, the louder one hides the quieter one: your ear genuinely cannot resolve the second while the first is playing. The same happens in time: a loud transient masks quiet detail immediately before and after it. The encoder models this and discards what is masked.
At a generous bitrate that model is conservative and the result is close to transparent. As you lower the bitrate, the encoder has to become more aggressive about what counts as “masked”, and eventually it starts discarding things you can hear. That is when artifacts appear: cymbals turning to hiss, reverb tails going grainy, a swirling quality on sustained notes.
Where the wheels come off: generation loss
Everything above describes one encode from a clean source. The failure mode that ruins real audio is doing it repeatedly.
An encoder does not know that its input is already damaged. Hand it a 128 kbps MP3 that has been decoded back to raw audio, and it analyses that signal on its own terms, artifacts included. It will faithfully spend bits preserving the previous encoder’s distortion, because to this encoder that distortion is simply part of the music.
Do that a few times and the losses compound in a way that is much worse than a single aggressive encode. This is the mechanism behind audio that has been forwarded through half a dozen apps and arrives sounding underwater.
The practical rule follows directly: keep an untouched original and always encode from it. If you are producing several versions of the same recording, generate each one from the source rather than from the last output. Our MP3 compressor is a good example of a tool it is tempting to run twice on the same file; running it once on the original and once more on the result is exactly the mistake.
The direction that does not work
Converting a lossy file to a lossless one is the most common misconception in this area, and it is worth stating flatly: it recovers nothing.
A lossless format is a perfect preservation mechanism. Give it a signal that has already had detail removed and it will preserve that signal perfectly, every remaining sample, every artifact, at several times the file size. MP3 to WAV is genuinely useful when you need uncompressed audio for editing software that will not accept compressed input, or for a device that only reads WAV. It is not useful for quality, and no tool anywhere can be, because the discarded information does not exist in the file.
The reverse direction is where the real decision sits. FLAC to MP3 starts from a lossless source, which is why its quality options top out at 320 kbps and its lowest setting is 192; there is no reason to be stingy when the input is pristine and the whole point is a one-time conversion you will keep. WAV to MP3 offers 320, 192 and 128 for the same reason, with 192 as the default.
Picking a bitrate you will not regret
Bitrate is best understood as insurance, and the premium depends on what you are protecting.
Music you care about, on decent headphones: 256 to 320 kbps. Dense, bright, transient-heavy material (cymbals, strings, applause, anything with a lot of high-frequency energy) is where lower settings show first.
Music for background listening, phone speakers, the car: 192 kbps is genuinely fine, and the environment will mask far more than the encoder did.
Speech: far lower, and this surprises people. A voice recording is a narrow-band signal from a single source with no stereo image worth keeping, so there is much less to encode. A lecture at 40 kbps in mono remains perfectly intelligible where music at that setting would be unlistenable. That asymmetry is why our compression presets drop to mono only for the voice-focused option, halving the channel count is free for speech and destructive for music.
Anything you might re-encode later: as high as you can, or keep the lossless original. This is the case people underestimate, because the cost only appears in a future you have not thought about yet.
The archive rule, and why it is the only decision that matters
Almost every regret in this area traces back to one thing: someone discarded the original.
If you keep a lossless archive, every conversion you ever do starts from a pristine source. Change your mind about bitrate, need a different format, want a mono version for a podcast feed, all of it regenerates cleanly from the original, indefinitely. Storage is cheap and getting cheaper.
If you keep only lossy copies, every future conversion is a generation loss, and the decision you made years ago about what “good enough” meant is now permanent. You cannot go back, and no amount of processing will reconstruct what was thrown away.
That is the whole argument, and it is why the interesting question is not “lossless or lossy”. It is “what am I keeping, and what am I making copies of”. Keep the original lossless. Make the copies lossy, from the original, every time.
