gstreamer/docs/random/uraeus/gstreamer_and_midi.txt

This document was created by editing together a series of emails and IRC logs.
This means that the language might seem a little weird at places, but it should
outline most of the thinking and design adding MIDI support to GStreamer so far.

Authors of this document include:

Steve Baker 		<steve@stevebaker.org>
Leif Johnson 		<leif@ambient.2y.net>
Andy Wingo 		<wingo@pobox.com>
Christian Schaller 	<Uraeus@gnome.org>

About MIDI
----------

MIDI (Musical Instrument Digital Interface) is used mainly as a communications
protocol for devices in a music studio. These devices could be physical entities
(e.g. synthesizers, sequencers, etc.) or purely logical (e.g. sequencers or
filter banks implemented as software applications).

The MIDI specification essentially consists of a list of MIDI messages that can
be passed among devices ; these messages (also referred to as "events") are
usually things like NoteOn (start playing a sound), NoteOff (stop playing a
sound), Clock (for synchronization), ProgramChange (for signaling an instrument
or program change), etc.

MIDI is different from other cross-device or inter-process streaming methods
(e.g. JACK and possibly some CORBA implementations) because MIDI messages are
discrete and usually only exchanged a few times per second ; the devices
involved are supposed to interpret the MIDI messages and produce sounds or
signals. The devices in a MIDI chain typically send their audio signals out on
separate (physical) cables or (logical) channels that have nothing to do with
the MIDI chain itself.

We want to have MIDI messages available in GStreamer pipelines because MIDI is a
common protocol in many existing studios, and MIDI is more or less a standard
for inter-device communications. With MIDI support in GStreamer we can look
forward to (a) controlling and being controlled by external devices like
keyboards and sequencers, and (b) synchronizing and communicating among multiple
applications on a studio computer.

GStreamer and MIDI
------------------

MIDI could be thought of in terms of dataflow as a sparse, non-constant flow of
bytes. GStreamer works best with near-constant data flow, so a MIDI stream would
probably have to consist mostly of filler events, sent at a constant tick rate.
It makes the most sense at this point to distribute MIDI events in a GStreamer
pipeline as a sequence of subclasses of GstEvent (GstMidiEvent or suchlike).

On-the-wire hardware MIDI connections run at a fixed data rate:

  The MIDI data stream is a unidirectional asynchronous bit stream at 31.25
  Kbits/sec. with 10 bits transmitted per message (a start bit, 8 data bits, and
  one stop bit).

Which is to say, 3125 bytes/sec. I would assume that the rawmidi interface would
already filter out the stop and start bits? dunno. How about the diagram on[1],
I found that to be useful. The MIDI specification is also available (though I
can't find it online at the moment ... might have to buy a copy), and there are
several tutorial and help pages (just google for MIDI tutorial).

There's another form of MIDI (the common usage?), "Standard MIDI files," which
essentially specify how to save and restore MIDI events in a file. We'll talk
about that in a bit.

[1] http://www.philrees.co.uk/#midi

MIDI and current Linux/Unix audio systems
-----------------------------------------

We don't know very much about the OSS MIDI interface; apparently there exists an
evil /dev/sequencer interface, and maybe a better /dev/midi* one. I only know
this from overhearing it from people. For latency reasons, the ALSA MIDI
interface will be much more solid than using these devices ; however, the
/dev/midi* devices might be more of a cross-platform standard.

ALSA has a couple ways to access MIDI devices. One way is the sequencer API.
There's a tutorial[1], and some example code[2] -- the paradigm is 'wait on some
event fd's until you get an event, then process the event'. Not very
GStreamer-like. This API timestamps the events, much like Standard MIDI files.

The other way to use MIDI with alsa is with the rawmidi interface. There is a
canonical reference[3] and example code, too[4]. This is much more like
GStreamer. I do wonder about the ability to connect to other sequencer clients,
though...

[1] http://www.suse.de/~mana/alsa090_howto.html#sect04
[2] http://www.suse.de/~mana/seqdemo.c
[3] http://www.alsa-project.org/alsa-doc/alsa-lib/rawmidi.html#rawmidi
[4] http://www.alsa-project.org/alsa-doc/alsa-lib/_2test_2rawmidi_8c-example.html#example_test_rawmidi

Getting MIDI into GStreamer
---------------------------

All buffers are timestamped, and MIDI buffers should be no exception. A buffer
with MIDI data will have a timestamp which says exactly when the data should be
played. In some cases this would mean a buffer contains just a couple of bytes
(eg, NoteOn). If this turns out to be inefficient we can deal with that later.

In addition to integrating more tightly with GStreamer audio pipelines (see the
dparams and midi2pcm sections below), there are several elements that we will
need for basic MIDI interaction in GStreamer. These basics include file parsing
and encoding (is that the opposite of parsing ?), and direct hardware input and
output. We'll also probably need a more generic sequencer interface for defining
elements that are capable of sending and receiving this type of nonlinear stream
information.

For these tasks, we need to define some MIME types, some general properties, and
some MIDI elements for GStreamer.

Types :

- MIDI being passed to/from a text file : audio/midi (This is in my midi.types
  file, associated with a .midi or .mid extension. It seems analogous to a .wav
  file, which contains audio/wav-type information.)

- MIDI in a pipeline : audio/x-gst-midi ?

Properties :

- tick rate : (default to 96.0, or something like that) This is measured in
  ticks per quarter note (or "pulses per quarter note" (ppqn) to be picky). We
  should use float for this so we can support nonstandard (fractional)
  tempos.

- tempo : (default to 120 bpm) This can be measured in bpm (beats per minute,
  the musician's viewpoint), or mpq (microseconds per quarter note, the unit
  used in a MIDI file[1]). Seems like we might want a query format for these
  different units ? Or maybe we should just use MPQ and leave it to apps to do
  conversions ?

Elements :

- midiparse : audio/midi -> audio/x-gst-midi

  This element converts MIDI information from a file into MIDI signals that can
  be passed around in GStreamer pipelines. This would parse so-called Standard
  MIDI files (and XML format MIDI files ?). Standard MIDI files are just
  timestamped MIDI data; they don't run at a constant bitrate, and for that
  reason you need this element.

  The timestamps that this element produces would be based on the tempo
  property, and the time deltas of the MIDI file data. If no data exists for a
  given tick, the element can just send a filler event.

  The element should support both globbing and streaming the file. Streaming it
  is the most GStreamerish way of handling it, but there are MIDI file formats
  which are by definition unstreamable, therefore a MIDI plugin needs to support
  streaming and globbing - and globbing might be easiest to implement first. The
  modplug plugin also reads an entire file before playing so its a valid
  technique.

- ossmidisink : audio/x-gst-midi -> hardware

  Could be added to the existing OSS plugin dir, sends MIDI data to the OSS MIDI
  sequencer device (/dev/midi). Makes extensive use of GstClock to send out data
  only when the buffer/event timestamp says it should. (Could instead use the
  raw MIDI device for clocking, doesn't matter which.)

- alsamidisink : audio/x-gst-midi -> ALSA rawmidi API

  Guess what this does. Don't know whether alsa's sequencer interface would be
  better than its rawmidi one. Probably rawmidi?

- ossmidisrc, alsamidisrc : hardware -> audio/x-gst-midi

  Real time midi input. This needs to be from the rawmidi APIs.

It seems like we could implement a class hierarchy around these elements. We
could use a GstMidiElement superclass, which would include the above properties
and contain utility functions for things like reading from the clock and
converting between different time measurement units. From this element we ought
to have GstMidiSource, GstMidiFilter, and GstMidiSink parent classes. (Maybe
that's overkill ?) Each of these MIDI elements listed above could then inherit
from an appropriate superclass.

We also need an interface (GstSequencer) to allow multiple implementations for
one of the most common MIDI tasks (duh, sequencing). The midisinks could
implement this interface, as could other soft-sequencer elements like
playondemand. The sequencer interface needs to be able to support MIDI
sequencing tasks, but it should support more generic sequencing concepts.

As you might have guessed, getting MIDI support into GStreamer is thus a matter
of (a) creating a series of elements that handle MIDI data, and (b) creating a
sort of MIDI library (like Timidity ?) that basically includes #defines for MIDI
message codes and stuff like that. This stuff should be coded in the gst-plugins
module, under gst-libs/gst/sequencer (for the interface) and
gst-libs/gst/audio/midi/ (for the defines and superclasses).

Of course, this is just the basics ... read on for the really gory future stuff.
:)

[1] http://www.borg.com/~jglatt/tech/midifile/ppqn.htm

Looking ahead
-------------

- MIDI to PCM

It would be nice to be able to transform MIDI (audio/midi) to audio
(audio/x-raw-{int|float}), which could be further processed in a GStreamer
pipeline. In effect this would be using GStreamer as some kind of softsynth.

The first way to do this would be to send MIDI data to softsynths and get audio
data out. There's a very, very nice way of doing this in ALSA (the sequencer
API). Timidity can already register itself as a sequencer client, as can
amSynth, AlsaModularSynth, SpiralSynth, etc... and these latter ones are *much*
more interesting. This is the proper, IMHO, way of doing things.

But, the other question is getting that data back for use by GStreamer. In that
sense a librafied Timidity would be useful, I guess... see the thing is that all
of these sequencer clients probably want to output to the sound card directly,
although they are configurable. In this, the musician's only hope is Jack. If
the synth is jacked up, we can get its output back into GStreamer. If not, oh
well, it's gone ...

- MIDI to dparams

Once we have MIDI streams, we can start doing fun things like writing a
midi2dparams element which would map midi data to control the dynamic parameters
of other elements, but lets not get ahead of ourselves.

Which gets back to MIDI. MIDI is a representation of control signals. So all you
need are elements to convert that representation to control signals. In
addition, you'd probably want something like SuperCollider's Voicer element --
see [1] for more information on that.

All of this is pretty specific to a synthesizer system, and rightly so :
multiple projects use it it could go in some kind of library or what-what but
otherwise it can stay in individial projects.

[1] http://www.audiosynth.com/schtmldocs/Help/Unit_Generators/Spawners/Voicer.help.html

On using dparams for MIDI
-------------------------

You might might want to look into using dparams if:

- you wanted your control parameters to change at a higher rate thanyour buffer
  rate (think zipper noise and sample-granularity-interpolation)
- you wanted a better way to store and represent control data than midifiles
- We wrote a linear interpolation time-aware dparam so that we could really
  demonstrate what they're good for.

It was always the intention for dparams to be able to send values to and get
values from pads. All we need is some simple elements to do the forwarding.

Possible inefficiency remedy : GstControlPad
--------------------------------------------

If it turns out that sending MIDI events spaced out with filler (blank) events
isn't efficient enough, we'll need to look into implementing something new ; for
now, though, we'll just try the simple approach and hope our CPUs are fast
enough. But read on for a little brainstorming.

It seems like GStreamer could benefit from a different subclass of GstPad,
something like GstControlPad. Pads of this type could contain control data like
parameters for oscillators/filters, MIDI events, text information for subtitles,
etc. The defining characteristic of this type of data is that it operates at a
much lower sample rate than the multimedia data that GStreamer currently
handles.I think that control data can be sent down existing pads without making
any changes.

GstControlPad instances could also contain a default value like Wingo has been
pondering, so apps wouldn't need to connect actual data to the pads if the
default value sufficed. There could also be some sweet integration with dparams,
it seems like.If you want a default value on a control pad, just make the source
element send the value when the state changes. Elements that have control pads
could also have standard GstPads, and I'd imagine there would need to be some
scheduler modifications to enable the lower processing demands of control pads.

An example : integrating amSynth[1]
-----------------------------------

We would want to be able to write amSynth as a plugin. This would require that
when the process function is called, we have a MIDI buffer as input, containing
how ever many MIDI events occurred in, say, 1/100 sec for example, and then we
generate an audio buffer of the same time duration...)

Maybe this will indicate the kind of problems to be faced. GStreamer has solved
this problem for audio/video syncing, so you should probably do it the same way.
The first task would be to make this pipeline work:

  filesrc location=foo.mid ! midiparse ! amSynth ! osssink

midiparse will take MIDI file data as an input, and produce timestamped MIDI
buffers as output. It could have a beats-per-minute property as mentioned above
to specify how the MIDI beat offsets are converted to timestamps.

An amSynth element should be a loop element. It would read MIDI buffers until it
has more than enough to produce audio for the duration of one audio buffer. It
knows it has enough MIDI buffers by looking at the timestamp. Because amSynth is
setting the timestamps on the audio buffers going out, a MIDI sink element would
know when to play them. Once this is working, a more challenging pipeline might
be :

  alsamidisrc ! amSynth ! alsasink

This would be a real-time pipeline : any MIDI input should instantly be
transformed into audio. You would have small audio buffers for lowlatency (64
samples seems to be typical).

This is a problem for amSynth because it can't sit there waiting for more MIDI
just in case there is more than one MIDI event per audio buffer. In this case
you could either :

- listen to the clock so you know when its time to output the buffer
- have some kind of real-time mode for amSynth which doesn't wait forMIDI events
  which may never come
- have alsamidisrc produce empty timestamped MIDI buffers so that amSynth knows
  that is time to spit out some audio.

[1] http://amsynthe.sourceforge.net/amSynth/index.html