The major written reference for this article is version 1.0 of the MIDI specification, published by the International MIDI Association, copyright 1983. There exists an expanded document. This document, which I have not seen, is simply an expansion of the 1.0 spec. to contain more explanatory material, and fill in some areas of hazy explanation. There are no radical departures from 1.0 in it. I have also heard of a "2.0" spec., but the IMA claims no such animal exists. In any event, backwards compatibility with the information I am presenting here should be maintained.
I will give constants in C syntax, ie. 0x for hexadecimal. If I refer to bits by number, I number them starting with 0 for the low order (1's place) bit.

Physical connector specification

The standard connectors used for MIDI are 5 pin DIN. Separate sockets are used for input and output, clearly marked on a given device. The spec. gives 50 feet as the maximum cable length. Cables are to be shielded twisted pair, with the shield connecting pin 2 at both ends. The pair is pins 4 and 5, pins 1 and 3 being unconnected:

A device may also be equipped with a "MIDI-THRU" socket which is used to pass the input of one device directly to output.
I think this arrangement shows some of the original conception of MIDI more as a way of allowing keyboardists to control multiple boxes than an instrument to computer interface. The "daisy-chain" arrangement probably has advantages for a performing musician who wants to play "stacked" synthesizers for a desired sound, and has to be able to set things up on the road.

Electrical specification

Asynchronous serial interface. The baud rate is 31.25 Kbaud (+/- 1%). There are 8 data bits, with 1 start bit and 1 stop bit, for 320 microseconds per serial byte.
MIDI is current loop, 5 mA. Logic 0 is current ON. The specification states that input is to be opto-isolated, and points out that Sharp PC-900 and HP 6N138 optoisolators are satisfactory devices. Rise and fall time for the optoisolator should be less than 2 microseconds.
The specification shows a little circuit diagram for the connections to a UART. I am not going to reproduce it here. There's not much to it - I think the important thing it shows is +5 volt connection to pin 4 of the MIDI out with pin 5 going to the UART, through 220 ohm load resistors. It also shows that you're supposed to connect to the "in" side of the UART through an optoisolator, and to the MIDI-THRU on the UART side of the isolator.
I'm not much of a hardware person, and don't really know what I'm talking about in paragraphs like the three above. I DO recognize that this is a "non-standard" specification, which won't work over serial ports intended for anything else. People who do know about such things seem to either have giggling or gagging fits when they see it, depending on their dispositions, saying things like "I haven't seen current loop since the days of the old teletypes". I also know the fast 31.25 Kbaud pushes the edge for clocking commonly available UART's.

Data format

For standard MIDI messages, there is a clear concept that one device is a "transmitter" or "master", and the other a "receiver" or "slave". Messages take the form of opcode bytes, followed by data bytes. Opcode bytes are commonly called "status" bytes, so we shall use this term.
very similar to handling a terminal via escape sequences. There aren't ACK's or other handshaking mechanisms in the protocol.
Status bytes are marked by bit 7 being 1. All data bytes must contain a 0 in bit 7, and thus lie in the range 0 - 127.
MIDI has a logical channel concept. There are 16 logical channels, encoded into bits 0 - 3 of the status bytes of messages for which a channel number is significant. Since bit 7 is taken over for marking the status byte, this leaves 3 opcode bits for message types with a logical channel. 7 of the possible 8 opcodes are used in this fashion, reserving the status bytes containing all 1's in the high nibble for "system" messages which don't have a channel number. The low order nibble in these remaining messages is really further opcode.
If you are interested in receiving MIDI input, look over the SYSTEM messages even if you wish to ignore them. Especially the "system exclusive" and "real time" messages. The real time messages may be legally inserted in the middle of other data, and you should be aware of them, even though many devices won't use them.

Voice messages

I will cover the message with channel numbers first. The opcode determines the number of data bytes for a single message (see "running status byte", below). The specification divides these into "voice" and "mode" messages. The "mode" messages are for control of the logical channels, and the control opcodes are piggybacked onto the data bytes for the "parameter" message. I will go into this after describing the "voice messages". These messages are:

Many explanations are necessary here:

• For all of these messages, a convention called the running status byte may be used. If the transmitter wishes to send another message of the same type on the same channel, thus the same status byte, the status byte need not be resent.
• Also, a "note on" message with a velocity of zero is to be synonymous with a "note off". Combined with the previous feature, this is intended to allow long strings of notes to be sent without repeating status bytes.
• From what I've seen, the "zero velocity note on" feature is very heavily used. My six-trak sends these, even though it sends status bytes on every note anyway. Roland stuff uses it.
• The pitch bytes of notes are simply number of half-steps, with middle C = 60.
• On keyboard synthesizers, this usually simply means which physical key corresponds, since the patch selection will change the actual pitch range of the keyboard. Most keyboards have one C key which is unmistakably in the middle of the keyboard. This is probably note 60.
• The velocity bytes for velocity sensing keyboards are supposed to represent a logarithmic scale. "advisable" in the words of the spec. Non-velocity sensing devices are supposed to send velocity 64.
• The pitch wheel value is an absolute setting, 0 - 0x3FFF. The 1.0 spec. says that the increment is determined by the receiver. 0x2000 is to correspond to a centered pitch wheel (unmodified notes)
• I believe standard scale steps are one of the things discussed in expansions. The six-trak pitch wheel is up/down about a third. I believe several makers have used this value, but I may be wrong.
• The "pressure" messages are for keyboards which sense the amount of pressure placed on an already depressed key, as opposed to velocity, which is how fast it is depressed or released. I'm not really certain of how "channel" pressure works. Yamaha is one maker that uses these messages, I know.
• Now, about those parameter messages. Instruments are so fundamentally different in the various controls they have that no attempt was made to define a standard set, like say 9 for "Filter Resonance". Instead, it was simply assumed that these messages allow you to set "controller" dials, whose purposes are left to the given device, except as noted below. The first data bytes correspond to these "controllers" as follows:

• The second data byte contains the seven bit setting for the controller. The switches have data byte 0 = OFF, 127 = ON with 1 - 126 undefined. If a controller only needs seven bits of resolution, it is supposed to use the most significant byte. If both are needed, the order is specified as most significant followed by least significant. With a 14 bit controller, it is to be legal to send only the least significant byte if the most significant doesn't need to be changed.
• This may of, course, wind up stretched a bit by a given manufacturer. The Six-Trak, for instance, uses only single byte values (LEFT justified within the 7 bits at that), and recognizes >32 parameters.
• Controller number 1 is standardized to be the modulation wheel.

Mode messages

These are messages with status bytes 0xb0 through 0xbf, and leading data bytes 122 - 127. In reality, these data bytes function as further opcode data for a group of messages which control the combination of voices and channels to be accepted by a receiver.
An important point is that there is an implicit "basic" channel over which a given device is to receive these messages. The receiver is to ignore mode messages over any other channels, no matter what mode it might be in. The basic channel for a given device may be fixed or set in some manner outside the scope of the MIDI standard.

The meaning of the values 122 through 127 is as follows:

124 - 127 also turn all notes off.

Local control refers to whether or not notes played on an instruments keyboard play on the instrument or not. With local control off, the host is still supposed to be able to read input data if desired, as well as sending notes to the instrument. Very much like "local echo" on a terminal, or "half duplex" vs. "full duplex".
The mode setting messages control what channels / how many voices the receiver recognizes. The "basic channel" must be kept in mind. "Omni" refers to the ability to receive voice messages on all channels. "Mono" and "Poly" refer to whether multiple voices are allowed. The rub is that the omni on/off state and the mono/poly state interact with each other. We will go over each of the four possible settings, called "modes" and given numbers in the specification:

The spec. states that a receiver may ignore any mode that it cannot honor, or switch to an alternate - "usually" mode 1. Receivers are supposed to default to mode 1 on power up. It is also stated that power up conditions are supposed to place a receiver in a state where it will only respond to note on / note off messages, requiring a setting of some sort to enable the other message types.
I think this shows the desire to "daisy-chain" devices for performance from a single master again. We can set a series of instruments to different basic channels, tie 'em together, and let them pass through the stuff they're not supposed to play to someone down the line.
This suffers greatly from lack of acknowledgement concerning modes and usable channels by a receiver. You basically have to know your device, what it can do, and what channels it can do it on.
I think most makers have used the "system exclusive" message (see below) to handle channels in a more sophisticated manner, as well as changing "basic channel" and enabling receipt of different message types under host control rather than by adjustment on the device alone.
The "parameters" may also be usurped by a manufacturer for mode control, since their purposes are undefined.
Another HUGE problem with the "daisy-chain" mental set of MIDI is that most devices ALWAYS shovel whatever they play to their MIDI outs, whether they got it from the keyboard or MIDI in. This means that you have to cope with the instrument echoing input back at you if you're trying to do an interactive session with the synthesizer. There is DRASTIC need for some MIDI flag which specifically means that only locally generated data is to go to MIDI out. From device to device there are ways of coping with this, none of them good.

System messages

The status bytes 0x80 - 0x8f do not have channel numbers in the lower nibble. These bytes are used as follows:

The status bytes 0xf8 - 0xff are the so-called "real-time" messages. I will discuss these after the accumulated notes concerning the first bunch.
Song position / song select are for control of sequencers. The song position is in beats, which are to be interpreted as every 6 MIDI clock pulses. These messages determine what is to be played upon receipt of a "start" real-time message (see below).
The "tune request" is a command to analog synthesizers to tune their oscillators.
The system exclusive message is intended for manufacturers to use to insert any specific messages they want to which apply to their own product. The following data bytes are all to be "data" bytes, that is they are all to be in the range 0 - 127. The system exclusive is to be terminated by the 0xf7 terminator byte. The first data byte is also supposed to be a "manufacturer's id", assigned by a MIDI standards committee. THE TERMINATOR BYTE IS OPTIONAL - a system exclusive may also be "terminated" by the status byte of the next message.

Yamaha, in particular, caused problems by not sending terminator bytes. As I understand it, the DX-7 sends a system exclusive at something like 80 msec. intervals when it has nothing better to do, just so you know it's still there, I guess. The messages aren't explicitly terminated, so if you want to handle the protocol (esp. in hardware), you should be aware that a DX-7 will leave you in "waiting for EOX" state a lot, and be sending data even when it isn't doing anything. This is all word of mouth, since I've never personally played with a DX-7.

Some MIDI ID's:

Note the USA / Europe / Japan grouping of codes. Also note that Sequential Circuits snarfed id number 1 - Sequential Circuits was one of the earliest participators in MIDI, some people claim its originator.
Two large makers missing from the original lineup were Casio and Oberheim. I know Oberheim is on the bandwagon now, and Casio also, I believe. Oberheim had their own protocol previous to MIDI, and when MIDI first came out they were reluctant to ? go along with it. I wonder what we'd be looking at if Oberheim had pushed their ideas and made them the standard. From what I understand they thought THEIRS was better, and kind of sulked for a while until the market forced them to go MIDI.
Nobody seems to care much about these ID numbers. I can only imagine them becoming useful if additions to the standard message set are placed into system exclusives, with the ID byte to let you know what added protocol is being used. Are any groups of manufacturers considering consolidating their efforts in a standard extension set via system exclusives?

Real time messages

This is the final group of status bytes, 0xf8 - 0xff. These bytes are reserved for messages which are called "real-time" messages because they are allowed to be sent ANYPLACE. This includes in between data bytes of other messages. A receiver is supposed to be able to receive and process (or ignore) these messages and resume collection of the remaining data bytes for the message which was in progress. Realtime messages do not affect the "running status byte" which might be in effect.

All of these messages have no data bytes following (or they could get interrupted themselves, obviously). The messages: