Getting Started
Welcome to Music Embedding! We will go through some basic concepts in this short tutorial.
Hint
Be sure you have Music Embedding installed. To install Music Embedding, please run pip install music_embedding.
Hint
In this tutorial, we will use Pypianoroll to convert MIDI files into pianorolls. To install Pypianoroll, please run pip install pypianoroll.
Hint
In the tutorial, we will use this MIDI file as an example.
First of all, let’s import the pypianoroll library and interval and embedder modules from music_embedding library.
from music_embedding import interval, embedder
import pypianoroll
Now, let’s read the example MIDI file into a Multitrack object.
multitrack = pypianoroll.read("music embedding example.mid")
Let’s create an embedder object and use it to get the melodic intervals of the first track.
embedder_object = embedder.embedder()
melodic_intervals = embedder_object.get_melodic_intervals_from_pianoroll(multitrack.tracks[0])
print(melodic_intervals.shape)
Here’s what we got:
(384, 4)
The MIDI file we used has 4 4-beat bars. Pypianoroll assigns 24 pixels to each beat, so total number of pixels is 4*4*24 = 384, which is the first dimension of our melodic intervals. Music Embedding implements intervals using 4-dimension vectors, which is the second dimension of melodic_intervals. These features are:
- interval_order: int
first to seventh
- interval_type: int
-2: dim, -1: min, 0: perfect, 1: Maj, 2: Aug
- octave_offsetint8
octave offset of a compound interval, 0 if interval is not compund
- is_descendingboolean
true if interval is descending
Let’s create an interval object and use it to see the first 30 intervals in melodic_intervals.
interval_object = interval.interval()
for i in range(30):
interval_object.set_specs_list(melodic_intervals[i])
print(interval_object)
The output is:
Silence
perfect 1st
perfect 1st
perfect 1st
perfect 1st
perfect 1st
perfect 1st
perfect 1st
perfect 1st
perfect 1st
perfect 1st
perfect 1st
perfect 1st
perfect 1st
perfect 1st
perfect 1st
perfect 1st
perfect 1st
perfect 1st
perfect 1st
perfect 1st
perfect 1st
perfect 1st
perfect 1st
Maj 2nd
perfect 1st
perfect 1st
perfect 1st
perfect 1st
perfect 1st
Melodic intervals are defined as the relatation between two consequtive notes. Since there is no note before the very first note, the very first note is shown as silence.
Hint
When converting melodic intervals to pianorolls, argument origin defines the first note and everything will be calculated based on that.
Since each note is stretched through multiple pianoroll pixels, when converted into intervals, each note appears as its true interval followed by multiple perfect 1st intervals. Let’s fix this by applying Run-Length Encoding:
melodic_intervals_rle = embedder_object.get_RLE_from_intervals(melodic_intervals)
print (melodic_intervals_rle.shape)
Here’s what we got:
(46, 5)
The first dimension is the compressed 384 that we had before. The second dimension contains the four features of an interval plus the fifth dimension which shows how many repeatitions are summerized into one. Let’s get a closer look:
print('Count \t interval')
print('-' * 17)
for i in range(melodic_intervals_rle.shape[0]):
interval_object.set_specs_list(melodic_intervals_rle[i, : -1])
print(f'{melodic_intervals_rle[i,-1]} \t\t {interval_object}')
The output is:
Count interval
-----------------
1 Silence
23 perfect 1st
1 Maj 2nd
23 perfect 1st
1 Maj 2nd
23 perfect 1st
1 min 2nd
23 perfect 1st
1 Maj 2nd
23 perfect 1st
1 Maj 2nd
23 perfect 1st
1 Maj 2nd
23 perfect 1st
1 min 2nd
22 perfect 1st
1 Silence
12 perfect 1st
1 Descending perfect 8th
11 perfect 1st
1 Maj 7th
23 perfect 1st
1 Descending Maj 2nd
11 perfect 1st
1 Descending Maj 6th
11 perfect 1st
1 perfect 5th
23 perfect 1st
1 Descending Maj 2nd
11 perfect 1st
1 Descending min 3rd
11 perfect 1st
1 Maj 2nd
23 perfect 1st
1 Descending Maj 2nd
11 perfect 1st
1 Descending perfect 5th
5 perfect 1st
1 Maj 2nd
5 perfect 1st
1 Maj 2nd
5 perfect 1st
1 min 3rd
5 perfect 1st
1 Descending Maj 2nd
11 perfect 1st