The MIDI specification, explained

What MIDI defines

MIDI, the Musical Instrument Digital Interface, was published in 1983 by a group of American and Japanese instrument makers. It is a way for instruments and computers to tell each other what to play. It never carries sound itself: only events like "this key went down this hard", "this knob moved", "start now" or "switch to the organ sound".

The MIDI 1.0 specification, and the documents added to it over the years, cover four things:

• A hardware link: the 5-pin DIN cable and the electrical signal on it.
• A message protocol: the bytes that make up notes, controllers, clock and the rest. This is what survives in USB, Bluetooth, network and software MIDI.
• A file format: Standard MIDI Files (.mid), for storing timed messages.
• Agreements on meaning: General MIDI's instrument list, controller numbers, tuning, time code, MPE.

The hardware: 5-pin DIN and a current loop

A classic MIDI port is a 5-pin 180° DIN socket. Each cable goes one way, from an OUT to an IN. A THRU port passes a copy of what arrives at IN, so devices can be daisy-chained.

3.5 mm TRS jacks

Small devices often use 3.5 mm jacks with an adapter. The standard wiring, called Type A (2018), is: tip to pin 5, ring to pin 4, sleeve to pin 2. Some older gear uses "Type B", with tip and ring swapped, so an adapter for one won't work with the other.

Status bytes and data bytes

MIDI messages are made of 8-bit bytes, and the top bit decides what a byte is:

• Status byte (80–FF, top bit 1): starts a message and says what kind it is. For channel messages, the high four bits are the message type and the low four bits the channel, 0–F for channels 1–16.
• Data byte (00–7F, top bit 0): a value from 0 to 127. Each message type has a fixed number of them.

So 93 40 7F is: 9 note on, 3 channel 4, key 40 (64, E4), velocity 7F (127). Because data bytes can't have the top bit set, a receiver can always tell where a new message starts, even if it joins halfway through.

Running status

If a message has the same status byte as the one before, the sender may leave it out and send only the data bytes. The receiver keeps using the last status. A run of notes on one channel then costs two bytes each instead of three. It's why many keyboards send note off as 9n kk 00: a note on with velocity 0 keeps the same status byte as the note ons.

• Running status only applies to channel messages (80–EF).
• System common messages and SysEx (F0–F7) cancel it: the next channel message must send its status.
• System real-time messages (F8–FF) don't affect it, and can appear between any two bytes, even in the middle of another message.

Channel voice messages

These carry the performance. n is the channel, 0–F.

Pitch bend joins its two data bytes into one 14-bit number: value = mm × 128 + ll, from 0 to 16,383. 8,192 (E0 00 40) is no bend. How far the ends go is set by the receiver, ±2 semitones by default, and can be changed with RPN 0.

Notes, velocity and pitch

Keys are numbered 0–127 in semitones. 60 is middle C and 69 is A440. The frequency of note n in equal temperament is 440 × 2(n − 69) / 12 Hz.

Octave names differ by maker: most software calls 60 C4 (scientific pitch, used on this site), while Yamaha and some DAWs call it C3. The note number is what's sent; the name is only a label.

Velocity is how fast the key went down, 1–127. How it maps to loudness is up to the instrument, and most let you choose a curve. A note on with velocity 0 is a note off.

Control change

Control change, Bn cc vv, sets controller cc (0–127) on channel n to value vv. Controllers 0–31 can be paired with 32–63 as their fine halves for 14-bit precision (send the MSB first). Switches like pedals read 0–63 as off and 64–127 as on.

"Undefined" controllers are free for any use, which is why controllers like 20–31 and 102–119 are popular for MIDI learn and custom mappings.

RPN and NRPN: parameters beyond 128

For settings that need more than a controller number, MIDI uses parameter numbers. You first select a parameter with two controllers, then set it with Data entry (CC 6, and CC 38 for the fine part):

• RPN (registered, standard meanings): CC 101 (MSB) and CC 100 (LSB).
• NRPN (non-registered, each maker's own): CC 99 (MSB) and CC 98 (LSB).

Setting the pitch bend range to ±12 semitones on channel 1, then closing the selection:

B0 65 00   RPN MSB 0
B0 64 00   RPN LSB 0      → pitch bend sensitivity
B0 06 0C   data entry 12  → 12 semitones
B0 26 00   data entry LSB → 0 cents
B0 65 7F   RPN MSB 127
B0 64 7F   RPN LSB 127    → null, nothing selected

Channel mode messages

Controllers 120–127 aren't controllers: they change how a channel behaves.

Omni and mono/poly combine into four modes: Mode 1 omni on, poly; Mode 2 omni on, mono; Mode 3 omni off, poly (the usual one); Mode 4 omni off, mono, used for guitar controllers with one channel per string. Messages 124–127 also act as All notes off.

System common messages

System messages have no channel: they're for every device on the cable.

System real-time messages

One byte each, and allowed anywhere, even between the bytes of another message, so timing is never held up.

At 120 BPM the clock ticks every 20.8 ms. Receivers that follow it count 6 ticks per 16th note and use song position pointer (F2) to find their place before a continue.

System exclusive

SysEx carries anything the rest of MIDI doesn't: patch dumps, firmware, device settings. It starts with F0, then a manufacturer ID, then any number of data bytes (top bit 0), and ends with F7.

• IDs are one byte (01 Sequential, 40 Kawai, 41 Roland, 42 Korg, 43 Yamaha…) or three bytes starting with 00.
• 7D is for non-commercial and educational use.
• 7E (non-real-time) and 7F (real-time) are Universal SysEx, defined by the specification. The byte after is a device ID, where 7F means "all devices".

Standard MIDI Files

A .mid file is a series of chunks. Each has a four-letter type and a 32-bit length (big-endian), then its contents. Readers skip chunk types they don't know.

The header chunk: MThd

Track chunks: MTrk

A track is a list of events, each with a delta time in front: the number of ticks since the previous event in the same track. Events are MIDI channel messages (with running status allowed), SysEx, or meta events, which hold information for the player rather than for an instrument. Every track ends with End of track.

Variable-length quantities

Delta times and lengths use 7 bits per byte. The top bit of each byte is 1 if another byte follows and 0 on the last one. Small numbers stay small; the maximum is four bytes.

Meta events

SysEx in a file is written as F0, a variable-length length, then the bytes after F0 including the closing F7. F7 with a length is an "escape" for sending raw bytes, or the next packet of a SysEx split in time.

A whole file, byte by byte

One middle C, a quarter note long, at 120 BPM. 41 bytes:

4D 54 68 64 00 00 00 06 00 00 00 01 00 60   MThd · format 0 · 1 track · 96 ticks per quarter
4D 54 72 6B 00 00 00 13                  MTrk · 19 bytes follow
00 FF 51 03 07 A1 20                    at 0:  tempo 500,000 µs = 120 BPM
00 90 3C 64                             at 0:  note on, C4, velocity 100
60 80 3C 40                             +96:   note off, C4
00 FF 2F 00                             end of track

Open any file in the MIDI file inspector to see it decoded like this. To turn ticks into seconds, divide by the division to get quarter notes, then multiply by the current tempo, which can change during the song.

General MIDI

MIDI itself doesn't say what program 1 sounds like. General MIDI (GM, 1991) does, so a song plays with the right instruments on any GM device: a fixed list of 128 programs, a drum kit on channel 10, and a minimum feature set.

A GM instrument must:

• Play 24 notes at once and listen on all 16 channels, each with its own program.
• Use channel 10 for drums, with the standard key map below.
• Respond to velocity, channel pressure, pitch bend (±2 semitones by default) and controllers 1, 7, 10, 11, 64, 121 and 123, plus RPN 0, 1 and 2.
• Treat note 60 as middle C.

The 128 programs come in 16 families of 8:

Programs are numbered 1–128 for people, but sent as 0–127: program 1, Acoustic Grand Piano, is Cn 00.

GM2 (1999) extended this with more controllers, more drum kits and bank-selected variations. Roland's GS and Yamaha's XG were earlier supersets of GM.

MIDI Time Code

MTC carries SMPTE time code (hours, minutes, seconds, frames) so MIDI gear can lock to tape and video. While running, the master sends quarter frame messages, F1 0nnn dddd: nnn says which piece of the time this is (frame, seconds, minutes and hours, low and high nibbles), and dddd is the value. Eight of them make one full time, spread over two frames. To jump, it sends a full frame SysEx instead. Frame rates are 24, 25, 29.97 drop-frame and 30 fps.

USB, Bluetooth, network and the browser

The messages stay the same; only the way they travel changes.

• USB MIDI (1999): each message travels in a 32-bit packet. The first byte holds a virtual cable number (up to 16 ports on one USB connection) and a code index saying how many bytes follow. Running status isn't used. USB is much faster than 31,250 baud.
• Bluetooth LE MIDI (2015): messages are grouped into small packets with a millisecond timestamp, which the receiver uses to restore the timing.
• RTP-MIDI (RFC 4695, 2006; updated as RFC 6295): MIDI over Ethernet and Wi-Fi, with recovery of lost packets. It's what macOS calls Network MIDI.
• Web MIDI: a browser API that lets web pages use MIDI devices, in Chrome, Edge, Opera and Firefox, but not Safari. Every tool on this site uses it.

MPE: MIDI Polyphonic Expression

Pitch bend and channel pressure act on a whole channel. To bend one note of a chord, MPE (adopted in 2018) puts each note on its own channel:

• A zone is a master channel plus member channels. The lower zone uses channel 1 as master and channels 2 upward as members; the upper zone uses channel 16 as master and works downward.
• The master channel carries things for the whole instrument, like the sustain pedal. Each new note goes to a free member channel.
• On the member channels, pitch bend is the note's own bend (±48 semitones by default), channel pressure is its pressure, and CC 74 is the up-and-down "slide" axis.
• Zones are set up with RPN 6, the MPE configuration message, sent on the master channel with the number of member channels.

MIDI 2.0

MIDI 2.0, adopted in 2020, is the first major update to MIDI. It keeps the same musical ideas (notes, controllers, channels, programs) and adds resolution, more channels and two-way conversation:

• Universal MIDI Packet (UMP): messages in 32-, 64-, 96- or 128-bit packets, in 16 groups of 16 channels. A UMP can carry MIDI 1.0 messages too.
• Higher resolution: 16-bit velocity, and 32-bit values for controllers, pressure and pitch bend.
• Per-note control: per-note pitch bend and per-note controllers, without MPE's channel juggling, and note attributes such as exact pitch.
• 16,384 registered and assignable controllers, each in one message instead of an RPN sequence.
• MIDI-CI (Capability Inquiry): devices discover each other, switch on profiles (agreed behaviour for, say, a drawbar organ) and exchange properties like patch lists as JSON.
• Jitter-reduction timestamps and endpoint discovery messages.

Compatibility was a rule from the start: the specification defines how to translate between 1.0 and 2.0, and a MIDI 2.0 device falls back to 1.0 when the other end doesn't answer. The 5-pin DIN port stays MIDI 1.0.

Status byte map

Every status byte, at a glance. Byte counts include the status byte.

Corrections and questions are welcome. MIDI is a trademark of the MIDI Manufacturers Association; the official specifications are free to download from the MIDI Association.