Midi DF

Midi DF is a Python module for converting midi data into DataFrames containing information on each chord in the midi file. DataFrames can be exported as JSON to facilitate use in external applications. Chords are stored as lists of midi pitches. Further reading on "dyads" and "interval classes", as well as the visualization technique that inspired this project, can be found here.

Examples using exported JSON data: After Effects (using AE's Expressions system, based on JavaScript), Processing (p5.js) (realtime and interactive)

Required libraries are pandas (0.20.3), numpy (1.13.3), and midi (0.2.3).

Demo

In short, midi.read_midifile() --> [tempo_map() -->] midi_to_df() --> condense_df() --> [merge_tracks() -->] output_json().

Dependencies

import json
import pandas as pd
import numpy as np
import midi
import copy

import midi_df as mdf

Load a midi file using the "midi" library

kiev = midi.read_midifile("kiev.mid")

Print the summary table

kiev_s = mdf.track_summary(kiev)
kiev_s

	Track Name	Channel	Note Events	Tempo Track	Instrument ID
0	The Great Gate of Kiev		0	tempo track	{}
1	Staff	0	2790		{0}
2	Staff-1	1	2032		{0}

We can see that Track 0 is the tempo track, and Tracks 1 & 2 have note data on channels 0 and 1, both with instrument 0 (Acoustic Piano in the GM standard).

Get the tempo map

Since Track 0 is the default for ch, we don't need to specify it in the parameters for tempo_df().

kiev_tempo = mdf.tempo_df(kiev)
kiev_tempo.head(10)

	Tick	uSec/tick	Time (uSec)	Time (sec)
0	0	2403.843750	0	0.000000
1	16128	2604.166667	38769192	38.769192
2	22272	2403.843750	54769192	54.769192
3	35328	2314.812500	86153776	86.153776
4	46950	2332.088542	113056527	113.056527
5	47052	2349.619792	113294400	113.294400
6	47155	2367.421875	113536411	113.536411
7	47257	2385.494792	113777888	113.777888
8	47360	2403.843750	114023594	114.023594
9	47462	2422.479167	114268786	114.268786

Create the first midi DataFrame (Track 1)

kiev_df1 = mdf.midi_to_df(kiev,1,kiev_tempo)
kiev_df1.head(15)

	dTicks	Tick	Pitch	On/Off	Playing	Time (s)	Beat
0	0	0	NaN	None	[]	0.000000	0.00
1	0	0	NaN	None	[]	0.000000	0.00
2	0	0	NaN	None	[]	0.000000	0.00
3	0	0	NaN	None	[]	0.000000	0.00
4	0	0	NaN	None	[]	0.000000	0.00
5	0	0	63.0	on	[63]	0.000000	0.00
6	0	0	67.0	on	[63, 67]	0.000000	0.00
7	0	0	70.0	on	[63, 67, 70]	0.000000	0.00
8	0	0	75.0	on	[63, 67, 70, 75]	0.000000	0.00
9	720	720	63.0	off	[67, 70, 75]	1.730767	3.75
10	0	720	67.0	off	[70, 75]	1.730767	3.75
11	0	720	70.0	off	[75]	1.730767	3.75
12	0	720	75.0	off	[]	1.730767	3.75
13	48	768	65.0	on	[65]	1.846152	4.00
14	0	768	70.0	on	[65, 70]	1.846152	4.00

Condense the DataFrame (Track 1)

kiev_c1 = mdf.condense_df(kiev_df1)
kiev_c1.head()

	Beat	Time (s)	Playing
8	0.00	0.000000	[63, 67, 70, 75]
12	3.75	1.730767	[]
16	4.00	1.846152	[65, 70, 74, 77]
20	7.75	3.576919	[]
24	8.00	3.692304	[67, 70, 75, 79]

Create and condense the DataFrame for Track 2 in one line

kiev_c2 = mdf.condense_df(mdf.midi_to_df(kiev,2,kiev_tempo))
kiev_c2.head()

	Beat	Time (s)	Playing
8	0.00	0.000000	[43, 46, 51, 55]
12	3.75	1.730767	[]
16	4.00	1.846152	[41, 46, 50, 53]
20	7.75	3.576919	[]
24	8.00	3.692304	[39, 43, 46, 51]

Merge the DataFrames for Track 1 and 2

kiev_m = mdf.merge_tracks([kiev_c1, kiev_c2],[1,2])
kiev_m.head()

	Beat	Time (s)	Playing 1	Playing 2	Playing
0	0.00	0.000000	[63, 67, 70, 75]	[43, 46, 51, 55]	[67, 70, 43, 75, 46, 51, 55, 63]
1	3.75	1.730767	[]	[]	[]
2	4.00	1.846152	[65, 70, 74, 77]	[41, 46, 50, 53]	[65, 70, 41, 74, 77, 46, 50, 53]
3	7.75	3.576919	[]	[]	[]
4	8.00	3.692304	[67, 70, 75, 79]	[39, 43, 46, 51]	[67, 70, 39, 75, 43, 46, 79, 51]

All main DataFrame steps in one line

mdf.merge_tracks([mdf.condense_df(mdf.midi_to_df(kiev, t, kiev_tempo)) for t in [1,2]]).head()

	Beat	Time (s)	Playing 1	Playing 2	Playing
0	0.00	0.000000	[63, 67, 70, 75]	[43, 46, 51, 55]	[67, 70, 43, 75, 46, 51, 55, 63]
1	3.75	1.730767	[]	[]	[]
2	4.00	1.846152	[65, 70, 74, 77]	[41, 46, 50, 53]	[65, 70, 41, 74, 77, 46, 50, 53]
3	7.75	3.576919	[]	[]	[]
4	8.00	3.692304	[67, 70, 75, 79]	[39, 43, 46, 51]	[67, 70, 39, 75, 43, 46, 79, 51]

Export to JSON

output_json(kiev_m, "kiev.json")

The resulting JSON file can be found here.

Calculate the maximum chord size

print("Maximum chord size:",mdf.max_notes(kiev_m))
print("Dyads in maximum chord:", mdf.max_dyads(kiev_m))

Maximum chord size: 11
Dyads in maximum chord: 55.0

Display the maximum count of each interval

mdf.max_dyad_counts(kiev_m)

0      14
1       9
2      12
3      15
4      16
5      15
6       4
sum    55
dtype: int64

Note that "sum" refers to the maximum of the sums of dyads (not the sum of the maxima), equal to the output of max_dyads().

Sample of Track Summary with more instruments

mdf.track_summary(fake_sample_file, recommend=True)

Recommended tracks: [1,2,3,4,6]

	Track Name	Channel	Note Events	Tempo Track	Instrument ID
0	A Song		0	tempo track	{}
1	Guitar 1	0	3234		{25}
2	Guitar 2	1	1847		{26, 29}
3	Keyboard - Square Wave	2	376		{80}
4	Organ	3	1429		{18}
5	Drum Kit	9	2308		{0}
6	Flute Solo	10	145		{73}

We can see that the author of this midi file has named their tracks by instrument. From here, we can determine which set of tracks we wish to use in our DataFrame. For example, if we only want to output guitars we would use Tracks 1 and 2, or if we wanted the whole band, we would copy the "Recommended tracks" list. Track 5 is a percussion track, set to midi channel 9, so we can assume it does not carry useful melodic information.