The Timbre in Popular Song (TiPS) Corpus
The Timbre in Popular Song (TiPS) Corpus
Interactive Project Report
Published: December 4, 2023
Authors
Nicole Biamonte (McGill University) [PI], Lindsey Reymore (Arizona State University), Ben Duinker (McGill University), Leigh VanHandel (University of British Columbia), Nicholas Shea (Arizona State University), Matthew Zeller (Musical Instrument Museum), Christopher William White (University of Massachusetts Amherst), Jeremy Tatar (McGill University), Jade Roth (McGill University), and Kelsey Lussier (McGill University)
Research Assistants (Annotators)
Hannah Benoit (McGill University), Holly Bergeron-Dumaine (University of British Columbia), Victoria Boerner (University of Toronto), Claire Brillon (University of British Columbia), Henri Colombat (McGill University), Hannah Davis-Abraham (University of Toronto), Jesse Diener-Bennet (McGill University), Yulia Draginda (McGill University), Phillip Elssner (McGill University), Dennis William Lee (University of Toronto), Kelsey Lussier (McGill University), Payton Mackwood (University of British Columbia), Tân Nazare (McGill University), Jade Roth (McGill University), and Jeremy Tatar (McGill University)
Introduction
Our project analyzes the understudied parameters of timbre and texture and their interactions with musical form in a new popular-music corpus, Timbre in Popular Song (TiPS). This study addresses two problems in popular-music scholarship: lack of research on timbre and texture, and underrepresentation of non-male and non-white artists in both popular music and its scholarship. We have developed and refined a comprehensive system for annotating timbre and texture in popular music, as well as a novel sampling method, the Anti-Discriminatory Alignment System (ADAS), that identifies and addresses imbalances in representation of artists’ gender, race, and ethnicity.
Timbre (tone color) and texture (musical layers) are important stylistic and structural parameters that are highly perceptually salient. In popular music, timbre is an important marker: listeners can identify genres and often specific artists in under a second based on timbre (Gjerdingen and Perrott 2008, Mace et al. 2012], with no time to perceive other musical features that unfold gradually. Popular music is well known for the wide timbral variety afforded by both electronic and acoustic instruments, as well as the performance tradition of using gritty-sounding timbres for expressive effect—e.g., distorted electric guitar, the raspy or growly quality of vocal fry, or the “dirty,” gravelly saxophone sound created by humming. Formal sections in popular music are usually more clearly defined than in classical music and are often marked by contrasts in timbre and texture. Yet despite the importance of these parameters in musical expression and structure, and their centrality to music listening, their functional roles in this repertoire have not been closely examined or theorized.
Another problem in popular-music research addressed by our study is the inequitable representation of artists. The comparative neglect of music by women, people of colour, and other marginalized groups, which has contributed to music theory’s endemic sexism (Hisama 2019, Maus 1993, McClary 1991) and deeply ingrained white racial frame (Ewell 2020 and 2023, Gopinath 2009, Kang 2009), has only recently begun to be addressed by scholars. In this project, we seek to counter biases against non-white and non-male artists by addressing their underrepresentation in research on popular music. Extant popular-music corpora such as the Rolling Stone corpus (De Clercq and Temperley 2011, Temperley and De Clercq 2013, 2017) and McGill Billboard corpus (Burgoyne 2011, White and Quinn 2018, White et al. 2022), the two popular-music corpora that have been most used by music analysts, are heavily biased toward white male artists. Our new corpus has more equitable representation: our novel sampling method for song selection (Shea et al. 2024) balances genre typicality with gender and racial diversity as well as chronological representation. This helps counter the marginalization of women artists and artists of colour in many popular-music genres and studies. Our new framework for anti-discriminatory sampling in music corpus studies modifies the methodologies of our discipline to help disrupt real-world biases.
The initial iteration of the TiPS corpus consists of 400 songs from the 1990s, comprising 100 songs each from four disparate genres with contrasting musical styles: country, pop, heavy metal, and hip hop. Since our focus is to make comparisons across genres, we limited this initial stage to songs from a single decade. We plan to expand the corpus, adding 3200 more songs from six decades and three more important genres--rhythm & blues/soul, rock, and electronic dance music, for a total of 3600 songs released between 1960 and 2019, encompassing many of the main genres of popular music. A planned third phase of the project will add the genres of punk, funk, and ska/reggae.
Song selection in TiPS balances genre typicality and chronological representation with considerations of diverse identities: our approach to song selection considers artists’ gender, race, and ethnicity, to help counter the systemic biases often propagated in music corpora and better reflect real-world diversity. Details related to timbre, texture, and form for each song are encoded and will be analyzed by genre to identify normative timbral and textural combinations, as well as typical differences among genres. Previous popular music corpora have been encoded primarily with respect to form and pitch (e.g., Burgoyne et al. 2011, De Clercq & Temperley 2011); our TiPS database is the first multi-genre corpus to be constructed for encoding timbral and textural information (the hip-hop corpus in Duinker & Martin 2017 includes annotations for texture and, to some degree, timbre).
Corpus Construction
Our Anti-Discriminatory Alignment System or ADAS derives a smaller “child” corpus from a larger “parent” one by applying a sampling method that is intentionally designed to counterbalance discriminatory forces. The child corpus includes more diverse artist identities and also may better represent the musical diversity within a style or genre (See Example 1). For each of the four genres in the current TiPS corpus (pop, hip hop, metal, and country), we compiled a parent corpus of 150–225 songs based on pre-existing sources such as Billboard charts, and encoded demographic variables of gender, race, and ethnicity for each artist. This information was used to create a child corpus for each genre. The child corpora each consist of 100 songs, balanced for artist representation (limited to 5 songs per artist or group) and even chronological distribution across the decade (10 songs per year, 1990–1999).
Example 1: A schematization of the Anti-Discriminatory Alignment System.
Parent corpora for each genre were derived from various sources, as no single source provides parallel lists across genres. The parent corpus for pop music was drawn from Billboard’s “Top Songs of the 90s” list, which includes 500 songs in a variety of genres. Song titles were filtered through iTunes to identify genre; 182 songs that iTunes labelled as “pop” constitute the pop parent corpus. The hip-hop parent corpus includes songs that were ranked #1 on Billboard’s “Hot Rap Songs” chart between 1990 and 1999, totalling 163 songs. The country parent corpus consists of the top 20 “Hot Country Songs” from the Billboard Year-End charts from 1990 to 1999, totalling 200 songs. Unlike the other three genres, there are no Billboard charts for heavy metal, so we derived the metal parent corpus of 225 songs from four lists based on critical acclaim: Hartmann 2018, Loudwire Staff 2020, Pasbani 2017, and Podoshen 2017. As these lists included songs that might be classified as hard rock or alternative rock, we developed a genre-filtering method like that used for the pop parent corpus, but with additional input from expert metal listeners.
Our demographic encoding system, developed from work by Shea (Shea 2022, Shea et al. 2024), records whether the lead artist or any band members belong to a marginalized racial, ethnic, or gender group, when explicit in print or online resources. We adopted the racial and ethnic categories used by The Music Coalition of the Annenberg Inclusion Initiative at the University of Southern California, which classifies these identities as underrepresented in the music industry: Black/African American, Hispanic/Latino, Asian, Native American/Alaska Native, Native Hawaiian/Pacific Islander, Middle Eastern, and Other/Mixed Race (Smith et al. 2018). With respect to gender, we considered any gender identity other than male as subject to marginalization. We also recorded a variable called “primary” that reflects the agency of marginalized artists within their ensembles. This variable is encoded as positive if either the title/lead member of the ensemble or more than half of the group’s members identify as racial, ethnic, or gender minorities. Student research assistants were trained in a 90-minute session to learn and practice the demographic encoding method described above, contextualized by critical discussion about issues of equity, diversity, and inclusion and the limitations of our approach. Following this training, they encoded demographic variables of gender, race, and ethnicity for each artist across the four parent corpora. For each entry, demographic encoders included their sources of information and commented on their decision-making processes as needed. Demographic information was cross-checked by team members in two ways: verifying a random sample of 35% of the encoded data for each genre/decade unit, and reviewing each entry to ensure coherence between the details in the comments and the data.
Our study aims to make our corpus more representative of the overall population of the United States during the chronological window represented by the corpus. We also opted to make this a relative rather than absolute goal: the minimum acceptable proportion for representation in each genre’s child corpus was reckoned based on the diversity within its parent corpus, with more initial diversity resulting in higher targets than corpora with less initial diversity. We therefore set target proportions for gender and for race/ethnicity by calculating the geometric mean—an averaging function designed for proportions—between the proportion in the parent corpus and the relevant US Census data for that time period.
Based on our demographic encoding, we calculated the proportions of marginalized artists in each parent corpus by gender and race/ethnicity. We then set target proportions for gender and race/ethnicity for the child corpora, to make them more representative of the overall population of the United States during the chronological window represented by the corpora. We also opted to make this a relative rather than absolute goal: the minimum acceptable proportion for representation in each genre’s child corpus was based on the diversity within its parent corpus, with more initial diversity resulting in higher targets than corpora with less initial diversity. We therefore set target proportions for gender and for race/ethnicity by calculating the geometric mean—an averaging function designed for proportions—between the proportion in the parent corpus and the relevant US Census data for that decade. In constructing the child corpora, we retained all songs by marginalized artists from the parent corpora. Next, in cases where the target proportion was below the population proportion, we supplemented the corpus via other sources to increase representation. Such adjustments for gender were made to the country, hip hop, and metal corpora, and adjustments for race/ethnicity were made to the country corpus. Example 2 shows differences in gender representation by the proportion between parent and child corpora for each genre; it also shows the same comparison for racial/ethnic representation. We recognize that there are multiple parameters of diversity, and any sampling method is necessarily imperfect. Our approach is not a solution to deeply ingrained issues of inequity, but our anti-discriminatory sampling framework represents a step toward more inclusive music scholarship.
Example 2: Proportions of marginalized artists in parent and child corpora compared to proportions in the U.S. population for 1990-1999 based on census data (horizontal line).
Song Annotation
Our theoretical approach to encoding form, texture, and timbre in popular songs is based on recent corpus-study methods in music scholarship that demonstrate large-scale patterns across time, genre, and culture, allowing for data analysis providing insight into these trends. Our categories of formal section types are based on De Clercq 2012, Summach 2012, and Temperley 2018, which are substantially in agreement: core song sections are verses, choruses, and bridges; ancillary song sections are intros, prechoruses, postchoruses, solos/instrumentals, outros, and links. Our analysis of texture is based on Allan F. Moore’s model of rock texture as comprising melody, harmonic filler, bass, and explicit beat layers (Moore 2012), with the addition of Megan Lavengood’s novelty layer (2017, 2020). Our encoding of timbre is entirely new, influenced by research in Reymore 2020, Zeller 2020, and White 2021.
Encoded metadata for each song includes the artist(s), song and album title, Billboard rank, release year, and a link to the source recording used. Information about timbre and texture is encoded by formal section; each row in the dataset corresponds to one formal section of a song, with variables listed in columns. Repeating sections (e.g., verse, chorus) are numbered in the order they occur—e.g., the first and second choruses are labelled “C1” and “C2,” respectively. Variables related to timbre and texture were recorded for each formal section, as shown in Example 3.
Example 3: Timbre and texture variables encoded in the TiPS corpus.
These variables were created and refined collaboratively by the authors. After compiling an initial set of proposed variables, several authors applied them to sample songs. We discussed the results, revising variables and definitions in response. We then completed a second round of encoding using the refined variables and a second round of collaborative revision. Next, research assistants (RAs) were trained to encode variables and practice song annotation in a 2-hour workshop. Each RA completed five songs, which were carefully reviewed by one of the authors; validation assignments were distributed so that each student received detailed feedback from five different team members, and revised their annotations as needed. We established a Slack workspace for the project, which allows research assistants and team members to be in communication with each other while working. This resulted in a third stage of collaborative refinements to the variables and definitions, after which the RAs continued encoding. Encoding for a sample song is shown in Example 4 below, and encoded data for the entire pop corpus is available in the Appendix.
Example 4: Sample annotation for Janet Jackson, “That’s the Way Love Goes” (1993).
After annotations for each genre/decade unit are completed, all will be cross-checked by other encoders. Analytical disagreements will be resolved by team members with expertise in that genre. Team members will cross-validate 35% of the annotated corpus and measure inter-analyst agreement with a kappa statistic to strengthen the analyses. Thus, all songs will be analyzed by at least two research assistants, and many will be additionally verified by a team member.
Project Status
Song encoding of the 400-song TiPS corpus is complete, and we are currently cross-checking the annotations. Team members will measure inter-analyst agreement with a kappa statistic to strengthen the analyses, and analytical disagreements will be resolved by team members with expertise in that genre. The pop corpus is presently available as a CSV file (see below), and the country, metal, and hip-hop corpora will be finished soon, all eventually becoming open-access datasets. This allows our team and future researchers to do statistical analysis using a variety of programs and programming languages, such as Excel, R, or Python.
Our first analytical goal is to define normative timbral, textural, and formal characteristics of each genre and identify the principal similarities and differences in such norms among genres. For example, we will identify which timbral vocal qualities (e.g., vocal fry, breathiness, growling, belting) are typical of each genre, and how the function of background vocals varies across genres. We will examine how timbral and textural characteristics vary with formal function, and how they interact with genre. Statistical analysis of our annotations can answer questions such as: which instruments are most commonly used in a given formal section type, and how does this vary from genre to genre? We will also address relationships among the 14 annotated variables, such as the relationships between instrumentation and texture. Our second analytical goal is to identify changes in timbral and textural norms over time, both within and across genres. For example, do certain instruments or instrument combinations become more or less common in different genres over time? How do trends in timbre and texture over time map across genres, and can we observe cross-fertilization of timbral and textural techniques across genres? Such changes may relate to cultural or technological shifts and may help us to better understand how such factors interact with trends in popular music. From these analyses, we will develop new theoretical models for timbre and texture in popular music and will adapt and disseminate our findings both as scholarly output (presentations, publications) and pedagogical resources (e.g., syllabi, lesson plans).
Conclusion
Our study offers a model for creating a more equitable corpus of popular songs using a new demographically adjusted sampling method, as well as a new encoding system for timbre and texture, which are treated as structural parameters for the first time in a corpus of popular songs. The next phase of our project will expand the repertoire beyond our initial four genres of pop, country, metal, and hip hop to add three more important genres: rhythm & blues, rock, and electronic dance music. We will also expand the corpus beyond the 1990s to encompass all decades in the sixty-year span between the 1960s and 2010s. For a more concise version of this information see Reymore et al. 2022, and for a more detailed discussion see Shea et al. 2024 (both linked in the appendix).
Acknowledgements
This research has been supported by funding from the Analysis, Creation, and Teaching of Orchestration (ACTOR) Project, the Social Sciences and Humanities Research Council of Canada (SSHRC), and the Fonds de recherche du Québec - Société et culture (FRQSC).
Appendix
Reymore, Lindsey et al. 2022 — Proceedings of the 2022 Music Encoding Conference
Shea, Nicholas et al., 2024 — Article, Music Theory Online (forthcoming)
Works Cited
Burgoyne, John Ashley. 2011. “Stochastic processes & database-driven musicology.” PhD dissertation, McGill University.
Burgoyne, John Ashley, Wild, Jonathan, & Fujinaga, Ichiro. 2011. “An expert ground truth set for audio chord recognition and music analysis.” In Proceedings of the 12th International Society for Music Information Retrieval Conference, 633–38. https://ismir2011.ismir.net/papers/OS8-1.pdf
De Clercq, Trevor. 2012. “Sections and successions in successful songs: a prototype approach to form in rock music.” PhD diss., Eastman School of Music.
De Clercq, Trevor, and Temperley, David. 2011. “A corpus analysis of rock harmony.” Popular Music 30/1: 47–70. https://www.cambridge.org/core/journals/popular-music/article/abs/corpus-analysis-of-rock-harmony/C5210A8EC985DDF170B53124F4464DA4
Duinker, Ben, and Martin, Denis. 2017. “In search of the golden age hip-hop sound (1986–1996).” Empirical Musicology Review 12/1-2: 80–100. https://emusicology.org/article/view/5410
Ewell, Philip. 2020. “Music theory and the white racial frame.” Music Theory Online 26.2. https://mtosmt.org/issues/mto.20.26.2/mto.20.26.2.ewell.html
Ewell, Philip. 2023. On Music Theory, and Making Music More Welcoming for Everyone. University of Michigan Press.
Gjerdingen, Robert O., and David Perrott. 2008. “Scanning the dial: the rapid recognition of music genres.” Journal of New Music Research 37/2: 93-100.
Gopinath, Sumanth. 2009. “Diversity, music theory, and the neoliberal academy.” GAMUT 2/1. https://trace.tennessee.edu/gamut/vol2/iss1/3.
Hartmann, Graham. 2018. “10 greatest metal songs of the 90s (year by year).” Loudwire. https://loudwire.com/10-greatest-metal-songs-1990s-year-by-year/
Hisama, Ellie. 2019. “Getting to count.” Music Theory Spectrum 43/2: 1-15.
Kang, Youyoung. 2009. “Diversifying music theory.” GAMUT 2/1. https://trace.tennessee.edu/gamut/vol2/iss1/4
Lavengood, Megan. 2017. “A new approach to the analysis of timbre.” PhD diss., City University of New York.
Lavengood, Megan. 2020. “The cultural significance of timbre analysis: a case study in 1980s pop music, texture, and narrative.” Music Theory Online 26.3. https://mtosmt.org/issues/mto.20.26.3/mto.20.26.3.lavengood.html
Loudwire Staff. 2020. “Top 90 hard rock + metal albums of the ‘90s.” Loudwire. https://loudwire.com/top-hard-rock-metal-albums-1990s/
Mace, Sandra T., Cynthia L. Wagoner, David J. Teachout, and Donald A. Hodges. 2012. “Genre identification of very brief musical excerpts.” Psychology of Music 40/1 (2012): 112-128.
Maus, Fred Everett. 1993. “Masculine discourse in music theory.” Perspectives of New Music 32/1: 264-293.
McClary, Susan. 1991. Feminine Endings: Music, Gender, and Sexuality. University of Minnesota Press.
Moore, Allan F. 2012. Song Means: Analyzing and Interpreting Recorded Song. Ashgate.
Pasbani, Robert. 2017. “These are the 100 best metal songs of the ’90s according to Metal Hammer.” Metal Injection. https://metalinjection.net/lists/these-are-the-100-best-metal-songs-of-the-90s-according-to-metal-hammer
Podoshen, Jeff. 2017. “Here’s an alternate ‘Best Metal Songs of the 90s’ list.” Metal Insider. https://metalinsider.net/lists/heres-an-alternate-best-metal-songs-of-the-90s-list
Reymore, Lindsey. 2020. “Empirical approaches to timbre semantics as a foundation for music analysis.” PhD diss., Ohio State University.
Reymore, Lindsey, Ben Duinker, Matthew Zeller, Leigh VanHandel, Nicholas Shea, Christopher William White, Jeremy Tatar, Jade Roth, and Nicole Biamonte. 2022. “Encoding and analyzing the timbre in popular song (TiPS) corpus.” Proceedings of the Music Encoding Conference 2022, ed. David Weigl and Jennifer Bain (forthcoming).
Shea, Nicholas. 2022. “A demographic sampling model and database for addressing racial, ethnic, and gender bias in popular-music empirical research.” Empirical Musicology Review 17/1: 49-58. Forthcoming.
Shea, Nicholas, Lindsey Reymore, Christopher William White, Ben Duinker, Leigh VanHandel, Matthew Zeller, and Nicole Biamonte. 2024. “Diversity in music corpus studies.” Music Theory Online 30.1. Forthcoming.
Smith, Stacy L., Choueiti, Marc, & Pieper, Katherine. 2018. “Inclusion in the recording studio? gender and race/ethnicity of artists, songwriters, and producers across 600 popular songs from 2012-2017.” Annenberg Inclusion Initiative. https://assets.uscannenberg.org/docs/inclusion-in-the-recording-studio.pdf
Summach, Jason. 2012. “Form in top-20 rock music, 1955-89.” PhD diss., Yale University.
Temperley, David. 2018. The Musical Language of Rock. Oxford University Press.
Temperley, David, and Trevor de Clercq. 2013. “Statistical analysis of harmony and melody in rock music.” Journal of New Music Research 42/3: 187-204.
Temperley, David and Trevor de Clercq. 2017. “Musical structure: melody and harmony in popular music.” In The Routledge Companion to Music Cognition, ed. Richard Ashley and Renee Timmers, 165-177. Routledge.
White, Christopher W. 2021. “Deployments of change and novelty in a corpus of popular music.” Proceedings of the Future Directions of Music Cognition International Conference.
White, Christopher W., and Ian Quinn. 2018. “Chord context and harmonic function in tonal music.” Music Theory Spectrum 40/2: 314-335.
White, Christopher W., Joe Pater, and Mara Breen. 2022. “A comparative analysis of melodic rhythm in two corpora of American popular music.” Journal of Mathematics and Music 16/2: 160-182.
Zeller, Matthew. 2020. “Planal analysis and the emancipation of timbre.” PhD diss., Duke University.