The MIDI file (.mid, .midi)

A MIDI file is sheet music for machines. It doesn't record sound; it records performance instructions: press this key at this moment, this hard, on this channel; switch channel 10 to drums; slow down here. A synthesizer, a sound card, a software instrument or a phone's built-in synth turns the instructions into sound when the file is played.

MIDI (Musical Instrument Digital Interface) began in 1983 as a way for electronic instruments to talk to each other over a cable. The Standard MIDI File format was added in 1988, so that sequences could be moved between different programs and devices. It has hardly changed since, and almost every music program can still read and write it.

Storing instructions instead of sound has big consequences:

• Files are tiny. A whole song is often a few kilobytes, where a WAV of it would be tens of megabytes and an MP3 a few.
• Everything stays editable. You can change a wrong note, the tempo, the key or an instrument after the fact, and notation software can turn the file into a score.
• The sound depends on the player. The same file played by a cheap phone synth, a Windows PC and a professional sample library sounds completely different (see General MIDI below).
• Vocals and real recordings can't be stored. A MIDI file can hold lyrics as text, but not a singer's voice.

A Standard MIDI File is a series of chunks, each a 4-character type, a 32-bit length (most significant byte first) and the data. There's one header chunk, MThd, then one or more track chunks, MTrk. Readers are expected to skip chunk types they don't know. The header's format field says how the tracks relate:

Time is counted in ticks. The header's division usually says how many ticks make a quarter note (values such as 96, 480 or 960), and tempo events say how many microseconds a quarter note lasts. A tick therefore has no fixed length: changing the tempo changes how long every following tick takes. If a file has no tempo or time signature events, the specification says to assume 120 beats per minute and 4/4.

A track is a list of events, each preceded by a delta time: the number of ticks since the previous event, so events that happen together have a delta of 0. Delta times and some lengths are written as variable-length quantities: 7 bits per byte, most significant first, with the top bit set on every byte except the last. Small numbers take one byte, and the largest allowed value, 0FFFFFFF, takes four.

There are three kinds of event:

Running status saves space: if a channel message has the same status byte as the previous one, the status byte can be left out, as in the anatomy above. Meta and SysEx events cancel running status, so the next channel message must start with its status byte again. This is also why many files end notes with a note on with velocity 0 instead of a note off: the specification treats the two the same, and staying on note-on status lets a whole passage use running status.

A program change only says "instrument number 0". What number 0 sounds like is up to the synthesizer. Early on, every manufacturer numbered its sounds differently, so a file made on one keyboard could play a piano part as a trumpet on another.

General MIDI (GM), published in 1991 by the MIDI Manufacturers Association and the Japan MIDI Standards Committee, fixed this. It assigns a named instrument to each of 128 program numbers (0 is Acoustic Grand Piano), reserves channel 10 for percussion, with each note number a different drum, and requires at least 24 voices at once. General MIDI 2 (1999) added more sounds and controllers; Roland's GS and Yamaha's XG were earlier supersets.

GM guarantees the right instruments, not the same sound. The piano on a phone, the one built into Windows and a 2 GB sampled piano are all program 0. Software synths load their instruments from a SoundFont (.sf2) or a similar sample library, so changing the SoundFont changes the sound of every MIDI file you play. If a file must sound exactly as the composer heard it, export it as audio.

Use MIDI to move music between programs, to edit or rearrange it, to drive hardware instruments, or to turn it into notation. Use MP3, WAV or another audio format to let people hear a finished piece.

• On a Mac: drag the file into a software instrument track in GarageBand; it appears as one or more tracks, and you can choose the instrument that plays each.
• As sheet music: MuseScore (free, Windows, macOS, Linux) opens .mid, .midi and .kar files and converts them to notation. The result needs tidying, because a MIDI file has timing, not note spelling or layout.
• With VLC: VLC plays MIDI through its FluidSynth module, which needs a SoundFont: in Preferences, show All settings, go to Input/Codecs › Audio codecs › FluidSynth, and choose an .sf2 file. On Windows, MIDI playback is available in VLC 3.0 and later.
• On Windows: Windows has a built-in General MIDI synthesizer with a sound set provided by Roland, which apps can play through. Which media app plays a double-clicked .mid has varied between Windows versions; VLC with a SoundFont is a dependable choice.
• In a browser: browsers don't play MIDI files. The Web MIDI API lets a page talk to connected MIDI devices, and web players use JavaScript synths, but an <audio> element won't play a .mid.

To turn a MIDI file into audio, play it through a synthesizer and record or export the result: in a DAW such as GarageBand, export the song as audio; on the command line, the free FluidSynth synthesizer renders MIDI files to WAV with a SoundFont.

To make or inspect one in code, the Python package mido (pip install mido) reads and writes Standard MIDI Files. This script writes the 64-byte file from the anatomy above:

import mido

mid = mido.MidiFile(type=0, ticks_per_beat=96)
track = mido.MidiTrack()
mid.tracks.append(track)
track.append(mido.MetaMessage("track_name", name="99helpers", time=0))
track.append(mido.MetaMessage("set_tempo", tempo=500000, time=0))  # 120 bpm
track.append(mido.Message("program_change", program=0, time=0))   # piano
track.append(mido.Message("note_on", note=60, velocity=100, time=0))
track.append(mido.Message("note_on", note=64, velocity=100, time=0))
track.append(mido.Message("note_off", note=60, velocity=0, time=192))
track.append(mido.Message("note_off", note=64, velocity=0, time=0))
mid.save("tiny.mid")  # mido adds the end-of-track event

• .rmi (RIFF MIDI): a Standard MIDI File wrapped in a RIFF chunk, the container also used by WAV. Microsoft documented it in 1991, and a 2000 MIDI Manufacturers Association note extended it to carry Downloadable Sounds (DLS) instruments with the music.
• .kar: karaoke files, which MuseScore and many players accept as MIDI files.
• XMF: the Extensible Music Format, a container for MIDI and sounds that the MIDI Manufacturers Association published in 2001, with a Mobile XMF variant for phones.

MIDI 2.0 was adopted by AMEI and the MIDI Association in January 2020. It sends messages as Universal MIDI Packets of 32 to 128 bits, with 256 channels in 16 groups. The Standard MIDI File itself didn't change: MIDI 2.0 added new file formats next to it.

The MIDI 2.0 counterpart of a format 0 file is the MIDI Clip File (specification M2-116-U, version 1.0, June 2023): one sequence of Universal MIDI Packets, starting with the 8 bytes SMF2CLIP. It uses the extension .midi2 and the media type audio/midi-clip, registered with IANA in March 2024. A companion MIDI Container File, comparable to format 1, holds one or more clips on several tracks. For exchanging music with most software today, use a Standard MIDI File.

• It plays silence. There's no synthesizer behind the player: VLC needs a SoundFont, and some systems have no default MIDI output. It may also be a format 1 file played by something that only plays the first (conductor) track.
• The instruments are wrong. The file has no program changes, or uses GS or XG bank selections the synth doesn't know, or was made for a specific keyboard rather than General MIDI.
• The drums are piano notes. The drum part isn't on channel 10, the General MIDI percussion channel.
• It's too fast or too slow. The tempo events are missing (players then assume 120 bpm) or were lost in conversion.
• A note keeps sounding. A note on has no matching note off (or note on with velocity 0), often in a file that was cut short. Stop playback, or fix the file in an editor.
• The imported score is messy. MIDI records timing in ticks, not note values, ties or enharmonic spelling. Adjust the quantization in the import settings, or clean the score up by hand.

A MIDI file holds musical events, not code. Two things are still worth knowing:

• Parser bugs. Like any binary format, MIDI has length fields a crafted file can lie about. VideoLAN removed MIDI playback from VLC on Windows in version 2.1.0 for security reasons, and restored it in 3.0. Keep players and libraries up to date.
• SysEx reaches your hardware. System exclusive messages in a file are sent to whatever device plays it, and can change that synthesizer's settings or stored sounds. Be careful playing untrusted files into a hardware instrument you've set up by hand.