Reading duplicate note-off events results in zero-duration notes

[bzamecnik]

When the input MIDI file contains note-off events for already inactive notes (with note-off events given before) pretty-midi still creates note events with zero duration. Writing such notes then results in spurious long-running notes that were not present originally.

Note that in this case there were two overlapping notes: note-on, note-on, note-off, note-off.

Example: bach-js_kunst_der_fuge_01_(c)unknown[2].mid (source: kunstderfuge.com)

If we zoom in the the start of the track 1 (first voice). Original looks like:

And written one:

Mido print_track() - original:

<meta message midi_port port=1 time=0>
<meta message track_name name='Contrapunctus I' time=0>
<message control_change channel=0 control=121 value=0 time=0>
<message note_on channel=0 note=69 velocity=88 time=1536>
<message note_on channel=0 note=74 velocity=88 time=192>
<message note_on channel=0 note=69 velocity=0 time=0>
<message note_on channel=0 note=72 velocity=88 time=192>
<message note_on channel=0 note=74 velocity=0 time=0>
<message note_on channel=0 note=69 velocity=88 time=192>
<message note_on channel=0 note=72 velocity=0 time=0>
<message note_on channel=0 note=68 velocity=88 time=192>
<message note_on channel=0 note=69 velocity=0 time=0>
<message note_on channel=0 note=69 velocity=88 time=192>
<message note_on channel=0 note=68 velocity=0 time=0>
<message note_on channel=0 note=71 velocity=88 time=96>
<message note_on channel=0 note=69 velocity=0 time=0>
<message note_on channel=0 note=72 velocity=88 time=96> <-- note-on
<message note_on channel=0 note=71 velocity=0 time=0>
<message note_on channel=0 note=72 velocity=88 time=192> <-- another note-on!
<message note_on channel=0 note=72 velocity=0 time=0> <-- proper note-off
<message note_on channel=0 note=74 velocity=88 time=48>
<message note_on channel=0 note=72 velocity=0 time=0> <-- duplicate note-off!
<message note_on channel=0 note=72 velocity=88 time=48>
<message note_on channel=0 note=74 velocity=0 time=0>
<message note_on channel=0 note=70 velocity=88 time=48>
<message note_on channel=0 note=72 velocity=0 time=0>
<message note_on channel=0 note=69 velocity=88 time=48>
<message note_on channel=0 note=70 velocity=0 time=0>
<message note_on channel=0 note=62 velocity=72 time=48>
<message note_on channel=0 note=69 velocity=0 time=0>
<message note_on channel=0 note=74 velocity=72 time=48>
<message note_on channel=0 note=62 velocity=0 time=0>
<message note_on channel=0 note=73 velocity=72 time=192>
<message note_on channel=0 note=74 velocity=0 time=0>

Mido print_track() - written:

written - bad:
<meta message track_name name='Contrapunctus I' time=0>
<message program_change channel=0 program=0 time=0>
<message control_change channel=0 control=121 value=0 time=0>
<message note_on channel=0 note=69 velocity=88 time=1536>
<message note_on channel=0 note=69 velocity=0 time=192>
<message note_on channel=0 note=74 velocity=88 time=0>
<message note_on channel=0 note=72 velocity=88 time=192>
<message note_on channel=0 note=74 velocity=0 time=0>
<message note_on channel=0 note=69 velocity=88 time=192>
<message note_on channel=0 note=72 velocity=0 time=0>
<message note_on channel=0 note=68 velocity=88 time=192>
<message note_on channel=0 note=69 velocity=0 time=0>
<message note_on channel=0 note=68 velocity=0 time=192>
<message note_on channel=0 note=69 velocity=88 time=0>
<message note_on channel=0 note=69 velocity=0 time=96>
<message note_on channel=0 note=71 velocity=88 time=0>
<message note_on channel=0 note=71 velocity=0 time=96>
<message note_on channel=0 note=72 velocity=88 time=0> <-- first note-on
<message note_on channel=0 note=72 velocity=0 time=192> <-- first note-off
<message note_on channel=0 note=72 velocity=0 time=0> <-- second note-off
<message note_on channel=0 note=72 velocity=88 time=0> <-- second note-on!
<message note_on channel=0 note=74 velocity=88 time=48>
<message note_on channel=0 note=72 velocity=88 time=48> <-- another note-on!
<message note_on channel=0 note=74 velocity=0 time=0>
<message note_on channel=0 note=70 velocity=88 time=48>
<message note_on channel=0 note=72 velocity=0 time=0> <-- note off
<message note_on channel=0 note=69 velocity=88 time=48>
<message note_on channel=0 note=70 velocity=0 time=0>
<message note_on channel=0 note=62 velocity=72 time=48>
<message note_on channel=0 note=69 velocity=0 time=0>
<message note_on channel=0 note=62 velocity=0 time=48>
<message note_on channel=0 note=74 velocity=72 time=0>
<message note_on channel=0 note=73 velocity=72 time=192>
<message note_on channel=0 note=74 velocity=0 time=0>

If we extract the input sequence of two notes of pitch 72:

<message note_on channel=0 note=72 velocity=88 time=0>
<message note_on channel=0 note=72 velocity=88 time=192>
<message note_on channel=0 note=72 velocity=0 time=0>
<message note_on channel=0 note=72 velocity=0 time=48>

it seems that the meaning was as follows (two consequent notes of length 192 and 48):

<message note_on channel=0 note=72 velocity=88 time=0>
<message note_on channel=0 note=72 velocity=0 time=192>
<message note_on channel=0 note=72 velocity=88 time=0>
<message note_on channel=0 note=72 velocity=0 time=48>

The only problem was that the note-off event preceded the note-on event, even though they occurred at the same time. Instead we created one note of duration 192, then one zero-duration note and one long note that was closed after encountering next note-off for that pitch.

We should fix two bugs here:

• when processing events at single time step, first close open notes, then open new notes
• ignore duplicate note-off events when reading MIDI
• remove zero-duration events in case they appear

[bzamecnik]

After fixing the piano roll read from the original MIDI seems to be fixed. Note that that rest is gone.

Full piece:

[craffel]

Thanks for debugging this.

When the input MIDI file contains note-off events for already inactive notes (with note-off events given before) pretty-midi still creates note events with zero duration. Writing such notes then results in spurious long-running notes that were not present originally.

While I can understand why we are creating zero-duration notes, and why this is bad, I don't understand where writing zero-duration notes results in spurious long-running notes.