An MP3 is thousands of tiny frames, each 26 milliseconds long.
There's no overall header that describes the file. Each frame starts with its own 4-byte label saying how it's encoded, then carries a slice of sound. That's why you can cut an MP3 anywhere, start a stream mid-way, and why a player can resume after a glitch. Click a frame to read its label.
The encoder uses a model of human hearing to decide what to discard. It doesn't cut evenly; it removes the parts other sounds would drown out anyway. These are the three main tricks, and what each saves.
The MP3 standard had nowhere to store the title or artist, so a separate format called ID3 was added on. Players skip it before reading frames. Here's what a typical block holds, and why a small song can have a surprisingly big file.
Drop an MP3 to walk it frame by frame: the ID3 tags at the front, the Xing or LAME header that makes seeking and gapless playback work, every frame header, and anything stuck on the end. The sample is a real MP3 built byte by byte in your browser, variable-bitrate silent frames with ID3v2, LAME and ID3v1 tags. Everything happens on your device.
11111111 11111011 10010000 01000100TIT2Six seconds of silenceTPE1FileFormerTALBTest tones · Vol. 1TDRC2026TSSELAME 3.100 -V2 (built in your browser)ID3v1“Six seconds of silence” · 128 bytes at the endIdentification, limits and variants, from ISO/IEC 11172-3 and 13818-3.
The two version bits in every frame header pick one of these. Lower sample rates exist for speech and low-bandwidth streams.
Converting from MP3 never restores what the encoder removed, and each page says so plainly.