How MP3 compression works, and which bitrate you can actually hear
A CD holds about 10 MB per minute of stereo music; an MP3 of the same minute is 1 MB and, at a decent bitrate, most people cannot tell them apart. The trick is not clever maths alone but a model of hearing: the encoder throws away sounds your ears would have discarded anyway. This guide explains that model, the bitrate scale, and what the MP3 converter is doing when it encodes on your device.
The idea: masking
Psychoacoustics found that a loud sound hides quieter sounds near it in frequency and in time — a cymbal crash masks a soft note a few hundred hertz away for a few tens of milliseconds. Absolute hearing thresholds also vary by frequency: we are most sensitive around 2–5 kHz and much less so at the extremes. MP3 (MPEG-1 Audio Layer III, 1993) uses a psychoacoustic model to compute, for each slice of audio, how much noise can be added at each frequency without being heard, then spends its limited bits only where they'll be noticed.
Inside the encoder
- Filter bank. The signal is split into 32 frequency bands, then each band into 18 finer lines with a cosine transform — 576 frequency values per frame of about 26 ms.
- Psychoacoustic analysis. A parallel analysis finds the masking threshold for the frame — how much quantisation noise is inaudible in each band.
- Quantisation. Each band is rounded as coarsely as the threshold allows. This is where bits are saved and information lost; it's the audio version of JPEG's step (How JPEG compression works — and what the quality slider really does).
- Huffman coding and framing. The rounded values are entropy-coded and packed into frames with headers; a "bit reservoir" lets hard frames borrow from easy ones.
Joint stereo encodes the mid and side signals separately, spending fewer bits on the side channel where little happens — a legitimate saving that becomes audible only at low bitrates.
What bitrate changes
Bitrate is the bit budget per second; lower means coarser rounding and a lower cut-off frequency (encoders drop the top end first). Listening tests broadly agree: 64 kbps is fine for speech and obviously degraded for music; 128 kbps is where most casual listeners stop noticing on most material; 192 kbps is transparent for the great majority of people on the great majority of tracks; 320 kbps is insurance for cymbals, applause and dense mixes, where artefacts ("swirling" high frequencies, pre-echo before sharp transients) persist longest. Variable bitrate (VBR) spends more on hard passages and less on easy ones and is better than constant bitrate at the same average size. The bitrate calculator turns any of these into file sizes.
Generation loss and transcoding
Encoding is lossy each time. An MP3 decoded and re-encoded loses again — the masked components already removed are re-analysed and more is cut — so a track that has been through three encodes at 128 kbps sounds worse than one at 96. Never transcode from one lossy format to another if you can avoid it; keep a lossless master (WAV, FLAC) and encode once from that. Converting MP3 to WAV cannot restore anything; it merely stops further loss.
AAC, Opus and when MP3 still wins
AAC (used by Apple, YouTube and most streaming) and Opus (used by WhatsApp, Discord and WebRTC) are newer designs with better transforms and stereo tools, and sound as good at roughly two-thirds of MP3's bitrate; Opus at 96 kbps rivals MP3 at 160. MP3's patents have expired, every device made since 1998 plays it, and there is no licensing to worry about, which is why it remains the safe deliverable for anything a stranger's device must open. For your own streaming or archiving use FLAC (lossless) and let the player transcode; Sample rate and bit depth explained: what 44.1 kHz and 16-bit actually mean covers the numbers behind the lossless file.
Sources and further reading
The claims in this guide rest on these references, which were checked when the guide was last updated. Spotted an error? The contact page says how to report it.