SUPRA MIDI files

This page describes the format for MIDI files available in the SUPRA archive. There are two types of MIDI files: (1) “raw” MIDI files, where each musical hole on the physical roll is represented by a note in the MIDI data, and (2) “expressive” MIDI files, where the individual holes have been merged into single notes (debridged), note durations have been extended slightly due to the width of the tracker bar, and expression controls on the left/right sides of the roll are interpreted as attack velocities for notes (as sustain/soft pedals controls added).

MIDI files can be accessed with three methods:

1. SUPRA Github repository: This repository allows for downloading all files, either via the git program, or via a link to a zip file found on the repository homepage.
2. SUPRA homepage: For each entry on the list of music on the SUPRA homepage, there are links to the "raw" and "expressive" MIDI files on the right side of the listings. Right click on the "Mraw" or "Mexp" links to save the MIDI file to your local computer.
3. PURL pages: the expressive MIDI file is available in the download section of the image viewer found on the PURL pages (bottom left corner of the image window).

Raw MIDI files

For every hole in the original roll, a MIDI note will be stored in the “raw” MIDI file. These holes may represent single notes, but typically need to be grouped into musical notes. The image on the right shows an analysis of the roll image. Each red/orange/yellow bounding box around a hole is represented by a MIDI note in the raw MIDI file; however, these five holes actually represent a sigle musical note. The raw MIDI files contain an estimation of these groupings by giving a loud note attack to the start of the expected musical note (marked with the yellow line in the image), and holes that sustain the note are given a quite note attack (holes that have the orange line at the top of their bounding box. Notice the difference between “holes” and “punches”: A “punch” is a circular hole generated by a circular stamp used to create holes in the roll paper. A “hole” can be a single punch (in which case it should be circular), or it can be an overlap of several punches. In the example image there are five “holes” that represent eight punches.

The individual holes can be merged with further processing into a single note according to the BRIDGE_FACTOR parameter (see metadata further below). The raw MIDI file is not intended for listing purposes, but rather as input for further processing to produce recuts (making copies of the original paper rolls). The image on the right illustrages a series of holes on a piano roll that represent a single note. Each of the boxes that bound a hole on the roll is converted into a separate MIDI note in the “raw” MIDI file.

Expression holes are given a quiet note attack for proof listening. Regular note holes are also played softly unless initial analysis by the image processing software identifies the hole as the start of a note, in which case the note hole is played loudly. Note that the expression is not interpreted/applied to the “raw” MIDI file.

There are five tracks in the “raw” MIDI file:

• Track 0: tempo settings for acceleration emulation and metadata parameters.
• Track 1: Note holes in the bass register, using MIDI channel 2.
• Track 2: Note holes in the treble register, using MIDI channel 3.
• Track 3: Expression holes for the bass register, using MIDI channel 4.
• Track 4: Expression holes for the treble register, using MIDI channel 5.

Timing

In both the “raw” and “expressive” MIDI files, the (initial) tempo of the roll is stored in the ticks-per-quarter-note (pulses-per-quarter-note) parameter of the MIDI file header. For most digitized rolls, the roll tempo will be this value divided by

1. For example if the tick-per-quarter-note value is 600, then this means that the roll tempo marking is “100”. Each MIDI tick is equivalent to one pixel row in the original image.

The tick values represent rows in the scanned image. Since most rolls are scanned at 300 DPI (or more precisely 300.25 DPI), then a roll tempo of 100, meaning 10.0 feet/minute, equals 10.0 * 12 * 300 = 36000 pixels rows per minute. For each second that is 36000 / 60 sec/min = 600 pixel rows per second, which is the setting used for the ticks-per-second in the MIDI file header. In other words, ticks in the MIDI file directly represent pixel lines in the scan of the MIDI file from which the MIDI file was digitized. DEA rolls will have a different multiplication factor other than 6 (but these have not been scanned yet).

Time 0 in the MIDI file is the position of the first automatically identified musical note hole in the image, which is at the image row given in the FIRST_HOLE metadata parameter. If the first “musical” hole is an expression hole, the final “expressive” MIDI file may have silence at the start of the MIDI file, since the expression holes are silenced or removed from the expressive MIDI files. Sometimes the first identified hole is a random tear in the MIDI file (and these are not manually corrected).

Acceleration emulation

Raw MIDI files will usually include acceleration emulation in the tempo track (track 0). When a roll is performed on a mechanical instrument, the winding mechanism pulling the roll around the take-up spool is rotating at a constant speed. As the paper builds up the diameter of the take-up spool, the speed at which the paper is moving will increase. This is estimated to be 0.22%/foot, so every 3600 ticks in the MIDI file, there will be an updated tempo value which is 0.22% greater than the previous tempo. The effect of tempo acceleration is needed for rolls greater than about 100 feet, since the end of the roll will otherwise sound like it is slowing down without the emulation.

Note that the tempo meta messages in the MIDI file are only used for acceleration emulation. The “tempo” of the roll is found in the MIDI header, in the ticks-per-quarter-note parameter. This allows for the acceleration emulation and the starting tempo to be set independently. This also allows for the tick values in the MIDI file to represent pixel rows in the original image.

Expressive MIDI files

The expressive MIDI files are generated from the “raw” MIDI file. There are several pre-processing steps: (1) remove the bridging from the raw holes by merging holes that are closer than the BRIDGE_FACTOR parameter in the metadata found in track 0; (2) tracker-bar holes have a height which will increase the effective length of holes after bridging has been removed. After these two steps have been done, the dynamics of each register are modeled and applied to the velocities of the sounding notes. Typically the expression tracks (tracks 3 and 4 will be removed from the final expressive MIDI file), so there are usually 3 MIDI tracks in the expressive MIDI files:

• Track 0: tempo indications and metadata parameters.
• Track 1: Note holes in the bass register, using MIDI channel 2.
• Track 2: Note holes in the treble register, using MIDI channel 3.

Expression MIDI file add additional metadata parameter that describe how the realization of the velocities was calculated (see the metadata section below for more information).

Metadata parameters

All SUPRA MIDI files contain metadata fields describing the MIDI file, such as the title, composer, and performer, as well as detailed parameters describing the analysis of the original roll image and interpretation of expression applied to the roll. The metadata is stored as text meta messages in the tempo track of the MIDI file. Each parameter is stored in a separate text meta message. The basic structure of an entry is:

@KEY: VALUE

In other words, the parameter key is prefixed by an @ character and is followed by a : character. The KEY does not contain spaces, but uses underscores to represent spaces in the parameter name. The value contains any number of characters, with accented letters in title/composer/performer names represented by hexadecimal unicode numeric entities, such as “&#xE9” for the letter “é” (this encoding is taken directly from the Stanford Libraries’ XML catkey entries for the rolls). The order of the parameters is not fixed, since processing of the MIDI file may rearrange the ordering.

Metadata parameters can be extracted SUPRA MIDI files using the rolltext program (to be described more later).

All metadata entries from all files in the SUPRA digital archive are available in this file in the SUPRA Github repository. The file can be converted into JSON data by using the ATON Javascript library. Additionally, you can view this metadata in the JavaScript console of your browser by typing the variable name ROLLINFO:

Basic metadata parameters

Image-analysis parameters

The following parameters are related to the initial process of extracting musical information from the scan of a piano roll. These parameters are useful to know when aligning the MIDI file data with the original image. Scanned images of rolls always present the bass register on the left. For red Welte-Mignon rolls, this will display the text in the label as upside down (180 degree rotation); for most other rolls, the label will be upside down and mirrored horizontally along a vertical axis.

The diagram on the right shows the locations of various parameter measurements. The length of the image is segmented in four locations: (1) PRELEADER_ROW is the pixel row of the start of the roll at the tip of the leader; (2) LEADER_ROW is the pixel row of the end of the leader (approximately); (3) FIRST_HOLE is the pixel row of the first musical hole on the roll (sometimes includes accidental holes); (4) LAST_HOLE is the pixel row of the end of the last musical hole in the image. The row length of two vertical sections of the row are also given: (1) MUSICAL_LENGTH is the number of pixel rows between the start of the first musical hole and the end of the last musical hole; (2) END_MARGIN is the pixel row length between the end of the last musical hole and the bottom of the image.

The two HARD_MARGIN parameters specify the width of the margin on the left and right sides of the image where the roll does not enter within the MUSICAL_LENGTH. The two MAX_DRIFT parameters gives the additional “soft margin” width inside of the hard margin, where the roll edge will oscillate within during the musical-length region.

Other parameters are mostly related to quality control of the physical roll and image, including parameters counting the number of tears on the edge of the roll, dust in the image margins, and counts of suspicious holes.

Expression parameters

The following parameters are added when the expression tracks are processed: