The Billboard Melodic Music Dataset (BiMMuDa)

1.1 Motivation

The primary motivation for compiling BiMMuDa was to create a high-quality dataset of Western pop melodies. Most large MIDI datasets consist only of classical music (Kong et al., 2020; Hawthorne et al., 2019; Dorfer et al., 2018; Müller et al., 2011) or video-game music (Donahue et al., 2018). Existing symbolic datasets of popular music are summarized in Table 1. The existing high-quality datasets, POP909 (Wang et al., 2020) and RWC (Goto et al., 2002) are of non-Western popular music. The most notable MIDI dataset for Western popular music is the Lakh MIDI dataset (Raffel, 2016), a corpus of over 175,000 MIDI files scraped from various websites. However, there are several issues with the Lakh MIDI dataset. Metadata is not consistently available for files, so filtering the dataset by genre, year, or any other criterion is impossible. The quality of MIDI transcription within these datasets is inconsistent, and MIDI files of popular songs likely to be familiar to many are heavily outnumbered by less-familiar music (Ji et al., 2020), which may be an issue if one is attempting to conduct analyses or train models that simulate the experience of the average Western listener. Thus, there is a general lack of high-quality symbolic datasets of Western popular music.

To overcome these issues, we compiled BiMMuDa, a dataset of more than 1,000 MIDI melodies from the Billboard top five singles of every year from 1950 to 2022. The MIDI files consist of accurate, manual transcriptions by highly experienced musicians, accompanied by precise metadata (see Section 4). The melodies are from the most popular songs of each year, so the dataset as a whole, or particular time ranges within it, can represent the lifetime melodic exposure of a Western individual. Because the scope of the dataset is well defined, it can easily be extended to create a larger dataset (e.g., by extending to the top ten rather than the top five, covering more years, extracting audio features) or abridged if necessary (e.g., by selecting MIDI files only from songs released during a particular decade). A more detailed discussion of the dataset’s possible applications is presented in Section 6.

1.2 The billboard charts

BiMMuDa is intended to be a compilation of melodies from the five most popular songs of each year, from 1950 to 2022. The Billboard year-end singles chart emerges as the most suitable measure of popularity for two reasons. Firstly, it is the only aggregate measure of song popularity that exists both today and as far back as 1950. Billboard magazine started publishing weekly song-popularity charts in 1940 based on record sales, airplay, and jukebox activity data; in 1946 they began aggregating this weekly data into annual, or year-end, charts (Trust, 2021). Chart positions on the weekly and year-end Billboard charts are still considered the standard measures of a song’s popularity in the United States today. Secondly, Billboard has adjusted its formula for determining chart positions several times over its history to reflect the changing ways music is and can be enjoyed. In the 2000s, their charts started incorporating data on paid digital downloads and streams, and in the 2010s they began including YouTube video streaming data (Molanphy, 2013). This makes the Billboard year-end chart a more comprehensive indicator of popularity than any singular measure of a song’s performance.

However, it must be noted that Billboard’s charts are limited as a measure of popularity. First, the charts consulted for this dataset track year-end popularity only in the United States, which may not be an adequate indicator of global popularity or popularity in other Western countries. Artists and record labels have found ways to manipulate the charts throughout the years (Andrews, 2018). Conversely, Billboard policy has kept some extremely popular songs from ranking highly on the year-end charts or even appearing on them at all. Before 1991, due to how chart positions were calculated, cumulative points for songs released towards the end of the year were often split over two year-ends, causing them to rank lower than they would have if all points were cumulated over a single year. In the 1990s, some songs achieved massive success but were not eligible for Billboard’s year-end singles chart because they were not available for purchase as singles. For example, in 1998 the Goo Goo Dolls’ “Iris” was number one for 18 weeks on Billboard’s Hot 100 Airplay chart but did not appear on the year-end singles chart. These issues must be considered. Nonetheless, Billboard’s singles charts are still the most widespread measure of a song’s popularity in US culture, and the top five year-end singles from each year can provide insight into trends in popular Western music. The list of singles represented in BiMMuDa can be found in the metadata.

2.1 Main melody identification

The vast majority of popular music consists of a salient or main melody accompanied by various supporting melodies, chords, and rhythms. “Main melody identification” in this article refers to the process of determining which instrument of those featured in a song is carrying the song’s main melody. This was the first step of the transcription process for each song in BiMMuDa. For over 95% of the songs, either the entirety of the main melody is sung by a lead vocalist, or a lead vocalist alternates with one or more featured vocalists. There are a few exceptions. Some top singles are orchestral pieces where the main melody is played by one or more instruments and often switches between instruments. For some songs, there is no main melody, and therefore no melody transcription or correction was performed. Lists of songs with non-vocal main melodies and no main melody are provided in the GitHub repository.

If a multitrack MIDI file for the song was available (see Section 2.2), main melody identification simply meant isolating the track corresponding to the lead vocalist’s part. If no MIDI file was available, main melody identification and transcription occurred simultaneously.

Main melody identification can be a source of error for several reasons. First, sometimes there are two equally salient melodies playing simultaneously (e.g., duets). The dataset is strictly monophonic, so in these cases the melody that began first was selected as the main melody. Often the end of the main melody of one section in a song overlaps with the beginning of the main melody of the next section. In these cases, the melodies had to be clipped to remove overlap. Finally, vocal harmonies can be very close in volume to the main melody, making the main melody difficult to isolate.

2.2 Melody transcription and correction

Many websites provide free MIDI files of pop songs. These were searched thoroughly before undergoing the transcription process to minimize the amount of manual transcription that would be required to complete the dataset. MIDI files were found for about 90% of the songs, but very few were usable for this project. Many were karaoke MIDIs, so the main melody was not present. Others were type 0 MIDI files in which all voices resided on a single track. It was found that manually transcribing the main melody was faster than attempting to isolate it from such files.

High-quality MIDI files were found for seven of the 371 songs in the dataset. The files and their sources can be found in the GitHub repository. In these files, each instrument or voice corresponded to a single MIDI track, so that those which corresponded to the main melody could be easily extracted. In such cases, after main melody identification it was only necessary to thoroughly check the melodies by comparing the audio of the song with the transcribed melody bar by bar and making corrections as needed. Spotify was used to find the audio for all tracks.

If no high-quality MIDI file for the song was available, the main melodies were manually transcribed via repeated listening of the song. All MIDI files were transcribed and checked using FL Studio 20 and MuseScore. The MIDI files were encoded using Program 0, which is usually mapped to a piano sound on most MIDI playback devices. All velocity values of MIDI notes are normalized to a “medium-loud” value of 85, about two-thirds of the maximum possible velocity value, 127. Minor differences between repetitions of the same section (e.g., repetitions of the verse or chorus; see Section 2.3) are generally disregarded for consistency and for respecting the nature of a symbolic transcription. However, the durations of transitions between melodic sections are accurately represented, as are tempo and key changes, which are annotated via the typical musical notation in the MuseScore files. The full transcriptions have not been aligned with the audio tracks.

2.3 Segmentation

Most popular songs consist of unique sections arranged according to a common format, or song structure. For example, many early-20th-century jazz and pop standards utilized “AABA” form: the first (usually 8-bar) section, or “A” section, is played and then repeated once before the “B” section and a final repetition of the A section (von Appen and Frei-Hauenschild, 2015). Other prevalent song structures in Western popular music include ABAB, AABB, and ABABCB (Watson, 2003). There are also common types of song sections. The A sections are often called verses, whose lyrics advance the narrative the song is conveying. Repetitions of verses usually feature different lyrics, though the melody through which the lyrics are sung typically stays consistent across repetitions (Davidson and Heartwood, 1996). A chorus typically follows the verse, and therefore is often the B section. Some songs have a pre-chorus between the verse and chorus. Choruses contain the central “idea” of the song, both lyrically and musically, and are usually more emotionally intense than verses (Davidson and Heartwood, 1996). There is typically no variation between repetitions of the chorus (Everett, 1999). Pre-choruses normally serve as harmonic transitions between the verses and choruses (Everett, 2009). A bridge usually refers to a section between two repetitions of a chorus that appears later in the song, which is meant to contrast with the verse section (Everett, 1999).

These song structures prevalent in Western popular music, and thus in the music represented in BiMMuDa, provide natural points of segmentation, and partitioning the melodic transcriptions in BiMMuDa using these points greatly expands the applicability of the dataset. Full melodic transcriptions of songs are lengthy, around the same length as the full audio track. While such transcriptions may be ideal for studies of song structure and experiments involving entire pieces of music, there are many cases where shorter melodies are preferable. Most behavioral studies on memory for melody, aesthetic responses to melody, and melodic expectation utilize stimuli less than a minute long (Pearce and Wiggins, 2006; Plantiga and Trainor, 2005; Schmuckler, 1989; Clemente et al., 2020; Gold et al., 2019). Therefore, in addition to providing full melodic transcriptions of the songs in BiMMuDa, it was decided that melodic transcriptions of the distinct sections of each song would also be copied and exported into separate files after the full transcription was created. For a song in BiMMuDa that follows, for example, an ABABCB structure, a full transcription of the main melody is provided, as well as three files containing transcriptions of the main melody for the A, B, and C sections.

Identifying song structure was mostly a straightfoward endeavor. Section boundaries are usually signaled with very noticeable changes in instrumentation and texture, so song structure usually became apparent during transcription. Lyrics websites such as Genius.com also often correctly segment lyrics according to song structure, so these were checked in cases where song structure did not become clear during transcription. Occasionally, segmentation was subjective: sometimes it was unclear whether one section differs enough from another to be considered unique, and if two short sections always occur together but are very distinct in some way (e.g., they feature different sets of instruments), it must be decided whether to deem them separate sections or combine them. However, with this dataset segmentation difficulties were minimal.

After transcription and correction, the melodies of each section were exported to separate MIDI files according to the identified song structure. The files are named according to the year, chart position, and order in which the section appears in the song. For example, “1965_01_3.mid” contains the main melody of the third section of the number one song in 1965, according to Billboard. The full transcription files contain a “full” suffix. The full transcriptions are encoded in both the MIDI and MuseScore formats.

2.4 Lyrics and metadata

If applicable, lyrics of each song were obtained from free lyrics websites and saved in a .txt file. The lyrics have not been aligned with the transcribed melodies. Metadata is available via two .csv files, one for per song attributes and one for per melody attributes, described in Tables 3 and 6, respectively. Summary statistics for both types of attributes are given in Section 4. Some attributes, such as the song titles, artists, and Spotify links, were created manually during the transcription process. Tunebat.com (https://tunebat.com/) provided intial values for the Tonic, Mode, and BPM attributes that were corrected manually. The remaining attributes were computed automatically. Section labels were assigned according to the guidelines of Gotham et al. (2021). Section labeling can be subjective, and some songs have N/A values.

3.1 Measuring compression distances between CoCoPops and BiMMuDa

An unbiased way to compare BiMMuDa and CoCoPops transcriptions is to calculate and examine a quantitative measure of similarity between them. The compression distance method developed by Li et al. (2004) and applied to music by Pearce and Müllensiefen (2017) is employed to this end as a perception-informed way of determining the similarity between two pieces of music. The measure is essentially the predictability of a piece of music according to a statistical model trained on a different piece. The higher the probability of the piece, the more similar the two pieces are, or, equivalently, the fewer operations one must do to transform one piece into the other. Because it closely simulates musical similarity perception between pieces (Pearce and Müllensiefen, 2017), the measure is an appropriate choice for this analysis.

Prior to analysis, some preprocessing of the CoCoPops transcriptions was needed to make them comparable to BiMMuDa transcriptions. While BiMMuDa transcriptions are strictly monophonic, the CoCoPops transcriptions include both supporting and main vocal melodies. Thus, the main melodies had to first be extracted from the CoCoPops files. The dissimilarity measure was computed between every possible pair of transcriptions where there is one transcription from each dataset, yielding values. These are visualized in matrix form in Figure 1. The units are in bits, and lower values indicate that fewer operations are needed to transform one piece into another, so lower values indicate higher similarity. The values can be divided into two groups: the non-diagonal values measure dissimilarity between transcriptions of two different songs, while the diagonal values measure dissimilarity between two transcriptions of the same song. The prominent diagonal of the matrix suggests that the transcription pairs representing the same song agree with each other. This is confirmed by a Student’s two-sample t-test, where the non-diagonal values () were found to be sampled from a different distribution than the diagonal values () with . Therefore, it can be concluded that, at least in the intersecting set of songs, BiMMuDa and CoCoPops transcriptions are perceptually similar.

Figure 1

3.2 Qualitative comparison

Quantitative analysis indicates that BiMMuDa and CoCoPops transcriptions do not perfectly agree but are highly similar. To establish the source of the disparities, the pairs of transcriptions were examined side by side while listening to the audio. This comparison made clear the general differences in transcription strategy between the two datasets, as well as some song-specific disagreements. One general difference concerns the utilization of song structure (see Section 2.3), specifically the repetition of sections. In BiMMuDa, small differences in the melody between repetitions of the same section (e.g., repetitions of the verse), are not captured: the transcription of the first occurrence of the section is simply copied as needed. In contrast, CoCoPops encodes these small variations, which leads to more disparity between the latter portions of the transcription pairs when sections are repeated. Another difference is that CoCoPops transcriptions are slightly more detailed in their representations of pitch, readily using accidentals when vocalists, for example, sing a note slightly flat or sharp. Meanwhile, BiMMuDa transcriptions are less sensitive to minor deviations from the key’s pitches, since its transcription strategy aims to reflect intended pitches rather than the performed ones. Finally, unlike BiMMuDa, CoCoPops does not explicitly encode swing timing, leading to significant rhythmic differences in the transcriptions of songs with swing.

Most of the discrepancies between BiMMuDa and CoCoPops transcriptions can be explained by the broad differences in the approach to transcription described above. The relevant dissimilarities for each pair of transcriptions, as well as the song-specific disagreements, are reported in Table 2. Overall, the two datasets mostly agree despite different approaches to transcription.

4.1 Attributes per song

Table 4 describes the attributes per song in BiMMuDa, and Table 5 summarizes their descriptive statistics. Their correlation matrix is visualized in Figure 2. The mean number of melodies per song, overall and per decade, is given in Figure 3. The distributions of the Mode and Tonic attributes are visualized in Figures 4 and 5, respectively. Tempo is visualized overall and per decade in Figure 6. Finally, means of the lyrical attributes (Number of Words, Number of Unique Words, Unique Word Ratio, and Number of Syllables) are visualized overall and per decade in Figure 7.

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Across the entire dataset, a song has about three melodies on average, and about two-thirds of the songs have between two and four melodies. The major mode is more common in every decade, although the difference in frequency between major and minor modes is reduced after the 1960s. The distribution of tonics is relatively flat, with songs in C and G being particularly popular in the 1960s. As for tempo, mean BPM cycles over time: faster songs () are more common in the 1960s and 2010s, while the prevalence of slower songs () peaks in the 1990s. The number of words, unique words, and syllables are doubled in the 2010s as compared with the 1950s, but the number of unique words relative to the total number of words decreases. The correlation matrix suggests that the number of parts is moderately correlated with the number of words and the number of syllables. In other words, songs with more melodic sections have more words and syllables. Number of Words is almost perfectly correlated with Number of Syllables. It is also positively correlated with Number of Unique Words but negatively correlated with Unique Word Ratio, suggesting that songs with more words tend to have a higher absolute number of unique words but a lower proportion of unique words. Finally, Number of Syllables and Number of Unique Words are positively correlated.

4.2 Attributes per melody

Table 6 defines the attributes per melody considered in the present study regarding BiMMuDa, and Table 7 shows their descriptive statistics. Figure 9 reports their pair-wise Pearson correlations. The development of Length and Number of Note Events is plotted in Figure 10, and the distribution of song sections is reported per decade and overall in Figure 11. The remaining attributes per melody are visualized in Figure 8.

Figure 8

Figure 9

Figure 10

Figure 11

Overall, BiMMuDa melodies are about 23 seconds long and contain about 49 note events. Melody length cycles approximately every two decades, with local maxima in the early 1950s, the mid-1970s, and the mid-1990s, and local minima in the mid-1960s, the 1980s, and the 2010s. In contrast, the mean Number of Note Events per year nearly doubles around the year 2000.

Figure 11 shows that verses are the most common type of melody in the dataset. The vast majority of songs contain both a verse and a chorus, but it is more likely for a song to have more than one unique verse than it is for a song to have more than one unique chorus, hence the higher prevalence of verse melodies. N/A values are assigned to melodies in songs that do not have the typical verse-chorus structure (e.g., orchestral pieces and ballads), which are most prevalent in the 1950s. Finally, melodies in the “Other” category, which include intros, outros, pre-choruses, post-choruses, hooks, and breaks, appear more often in more recent songs. We analyze this attribute in more detail in the next section.