chorrrds: Music Chords Extraction

title: 'chorrrds: An R package for music chords extraction and analysis' tags: - R - chords - music information retrieval authors: - name: Bruna Wundervald^[Maynooth University] orcid: 0000-0001-8163-220X affiliation: "1" affiliations: - name: Bruna Wundervald index: 1 date: 26 July 2020 bibliography: paper.bib

Summary

The Music Information Retrieval research field, in general terms, is concerned with the extraction and inference of relevant features from music. That can involve a big range of tasks, including music recommendation systems, analysis of music similarity and auto-tagging, automatic music recognition, artificial music composition, the relationship between music and the human behavior, and many more. Though the field is still young, it has been experiencing a constant growth in the past few years [@schedl2014music]. Considering that music is one of the most important cultural elements of our society and the impact it can have in our daily lives, this expansion is not surprising. Along with that, many software options for the analysis and creating tools involving music data are arising, such as ([@mcfee2015librosa], [@cuthbert2010music21], [@bock2016madmom], [@manaris2014making], [@giannakopoulos2015pyaudioanalysis]), to name a few.

However, given how broad MIR is, we are not even close to have all the possible tools and data that might be needed to perform specific tasks. Currently, MIR is mostly limited to the use of python [@van2007python], as can be perceived in the previous references to the software available. This constraint raises an issue regarding the accessibility and democratization of the field, which gets compromised by the fact that it demands knowledge in one very particular programming language. In addition, many MIR tasks are heavily related to statistics and machine learning, which, by their turn, hardly only focus in one primary programming language.

Another issue is the predominant focus on audio data. It is true that, with the increase of computation power, being able to process the signal straight from audio has been responsible for major improvements to the MIR research field, since audio data is exceptionally rich and informative. However, obtaining such music files can be very expensive due to rights-holders and copyright laws [@schedl2014music]. This has been a vital motivation for the difficult task of creating publicly available datasets of audio and audio-related features, and also one of the main problems that the researchers in MIR face. A less painful option is the use of symbolic data [@billard2006symbolic], which is comprised of digital representations of musical scores (MIDI, for example), with notes, timbre and pitch information, chords data, lyrics, and so on.

The existence of these two issues are the main motivation for the software proposed in this paper. The chorrrds package is mostly dedicated to the extraction of music chords in R [@rmanual]. Along with the main functions, the package is is accompanied by useful functions and objects to help analyze the music chords and inspire users in their research problems. The music chords are obtained via the web scraping [@webscraping] of the Cifraclub website, an online collaborative Brazilian page of music-sharing. The website freely provides a big collection of music chords for different instruments, and most of the chords information present there are contributions of the users. The chords for any artist or band available on the website can be extracted, as well as the results for the other functions of the package.

In the following, we exemplify the usage of the main and the auxiliary functions and object of the package, as well as provide a few practical usage examples.

Main functions

The chorrrds package is widely available through CRAN, the biggest repository of R packages. This makes its installation as simple as

R> install.packages("chorrrds")

that will retrieve and install the latest version available on CRAN. After that, we can load and use the get_chords() function, that depends on the results of the get_songs() function. This is easily done with, for example

R> library(chorrrds)
R> artist <- "janis-joplin"
R> song_urls <- get_songs(artist)
R> song_chords <- get_chords(song_urls)

where the song_urls and the song_chords objects are both a tibble. The last is composed by 4 data columns: chord, key, song and artist*, which represent, the chords collected (in temporal, but not necessarily equally spaced, order), the song key as found on the website, the name of the song and the name of the artist.

We recommend the use of the clean() function always, that removes strange objects if they end up being retrieved along with the chords, such as lyrics or other html elements. This rarely happens, but if it does, it is better to prevent that it gets mixed with the chords data, with

R> song_chords <- clean(song_chords)

Accompanying the main chords extraction function, we also provided a few other functions to help analyze and wrangle the extracted data. The first one is the chords_ngram(), that allows for the creation of n-grams [@damashek1995gauging] for the sequences of chords. This is useful in the case where the taks involves the patterns on the chords transitions present in the data, which is frequent in harmony analysis. Chords transitions are informative about the underlying harmonic structure present in songs, and it might serve for comparing songs, analyzing selected pieces, or even a full artist collection. The code for finding the n-grams is

R> jj_ngrams <- chorrrds::chords_ngram(song_chords, n = 2)
R> jj_ngrams[2, "chords_ngram"]
R> [1] "Dadd9 Dsus4/9"

The choice of n, or how many chords should be considered in each n-gram, will depend on the analysis being conducted, and is set to 2 as default.

Going further on the analysis tools provided, we have the feature_extraction() function, that calculates summary features from the chords dataset, indicating if each chord is:

minor,
diminished,
augmented,
sus,
chords with the 7th,
chords with the major 7th,
chords with the 6th,
chords with the 4th,
chords with the augmented 5th,
chords with the diminished 5th,
chords with the 9th, chords with varying bass

Again, those are variables that are constructed in order to facilitate the harmonic structure analysis of the pieces. For instance, one can use the given variables to analyze the harmonic complexity of the songs regarding many different levels. As for the code, it is only

R> jj_features <- feature_extraction(song_chords)
R> jj_features[1, c(1, 6:17)]
R> chord minor dimi augm sus seventh seventh_M sixth fourth fifth_aug fifth_dim ninth bass
R> Dadd9     0    0    0   0       0         0     0      0         0         0     1    0

Added to the set of analysis functions, we put some auxiliary objects in the package. For these cases, the choice of keeping them as objects instead of functions that use such objects is to give the user more freedom about the operations they can perform using the objects. The deg_maj and deg_min objects are basically dataframes composed by the chords in all the 12 major and minor scales. To give a brief explanation, a major scale is be defined by having the W–W–H–W–W–W–H note interval pattern (W = whole step, 2 semitones; H = half step, 1 semitone), starting in a certain note (as C, D, etc), and we can arguibly say that this is the most important scale structure in occidental music. The minor scale is defined by the W-H-W-W-H-W-W structure, and it follows the major scale on the importance ranking. Similarly, its chords will be the ones starting in each of its notes, but now Given that, the two mentioned objects can help the user to transpose songs to different scales, since when we collect the chords, we also collect the key in which the chords are. Both objects also contain a degree column, indicating which is the degree (I-II-III-IV-V-IV-V) of each chord in the scales. Coming back to harmonic analysis, much of what happens in music structures are intrinsically related to the degrees of the chords used, as each degree plays a different role in the music structure. For example, some chords (or degrees), when followed by a specific one or another degree, can give the song a feeling of tension, or rest, depending on what the intention of the composer is. To use the objects, it is only a matter of calling it. Another interesting object present on the package is the dist dataframe, composed by the distances of each note by semitones and by steps in circle-of-fifths.

One of the most recent features added to the package is the possibility of extracting lyrics and aligning them with the music chords. This is done by running, for example:

lyrics_chords <- create_dat(artist = artist, track = song_urls$song[1])
lyrics_chords <- create_net(lyrics_chords)
head(lyrics_chords)

R> chord lyric                           
R> G     will soon grow tired of waiting,
R> C     She’ll do crazy                 
R> G/    things                          
R> B     yeah,                           
R> D     on lonely                       
R> G     for the new men now and again

where the chord column represents the chord played, and the lyric column represents the words being sung at each part of the song. The new function opens up a whole new set of applications for the package, once the relationship between chords/harmony and music lyrics is one of the interests of the MIR community.

Example Analysis

This section is dedicated to giving a few concrete examples of what can be done using the chorrrds package.

In Figure 1, we see the 15 most frequent chords in Janis Joplin's songs and the correspondent densities for the features extracted from her songs. From these plots, we can have a better idea of the harmonic anatomy of the songs. For instance, the two plots are evidencing that the songs have mostly simple chords, with no alterations (such as added notes or a different chord in the bass), and the minor chords can be quite predominant. Besides that, some chords are clearly preferred by the artist, as their frequencies are much higher than the less common chords.

The 15 most frequent chords in Janis Joplin's songs and densities of the proportion of each extracted feature in the songs. { width=70% }

As for Figure 2, continuing on the harmonic pattern exploration, we see 15 songs with the highest number of different chords. Though a few chords have a high frequency, we still have songs with a considerable number of different chords. With Figure 1, it could be argued that Janis Joplin's songs are harmonically simple, given their limitation to a specific group of chords. Figure 2, on the other hand, comes with an opposite idea, showing that even having preferences by some chords, the artist managed to create songs with a big variation of chords, which is not an easy composition task. Of course, there are many other details of the songs that are not being considered here, but that is already a big piece of information.

The 15 songs with the highest number of different chords. { width=30% }

In Figure 3, on the other hand, we present the results of a Gaussian Mixture clustering analysis on the average of the features extracted by the feature_extraction() function. So the input of the clustering is the proportion of each feature extracted from the chords (minor, diminished, etc) for each Janis Joplin song. With that, we try to verify if there exists a reasonable way of clustering her songs using the extracted features. The plot shows us the average of each of those features in each resulting cluster. In cluster 1, for example, the average proportion of the number of chords with the 4th note, the 6th note and the 9th note are higher than in the other 2 clusters, implying that those 3 features might serve well to characterize a portion of the songs. As for cluster 2, its big difference compared to the other clusters is that its songs have a high proportion of minor chords, which is usually associated with more melancholic/sad songs, also a perfectly plausible classification. Cluster 3 relates to the most "basic" songs, that are limited to using simpler versions of the chords.

Cluster results. { width=50% }