- The paper argues that Music Information Retrieval should analyze Pop music through audio, lyrics, images, videos, reviews, and metadata rather than treating sound as the sole object of study.
- The paper reviews large-scale findings including declining lyric complexity, changing musical variety, rising loudness, evolving vocal patterns, and genre-specific cultural themes, while noting conflicting results and questioned statistical methods.
- The paper proposes three research agendas: mapping lyric topics, building an iconography of album covers, and tracking retro cycles through joint multimodal embeddings, while emphasizing corpus bias and external-validity challenges.
This survey article by Flexer argues that Music Information Retrieval (MIR) should study Pop music as a multi-modal socio-cultural phenomenon rather than as audio alone, reviews existing large-scale computational work on Pop music, and identifies two structural weaknesses in the field—scarcity of genuinely multi-modal studies and questionable external validity of corpus construction—before proposing three concrete research agendas (2608.17812).
Motivation and framing
The paper adopts a constructive definition of "Pop" as any music a person has been exposed to via mass media. A motivating example—Lee Perry's "Kung Fu Meets the Dragon" (1975) and the Wu-Tang Clan's "Enter the Wu-Tang (36 Chambers)" (1993)—illustrates that Pop music is only partly music: it interweaves audio with album artwork, typography, film references, production techniques, and cultural lineage. The author therefore proposes an approach analogous to "distant reading" in digital humanities (Moretti), extended to "distant listening," "distant viewing," and "distant reading" of lyrics, reviews, magazines, and metadata.
State of the art
Distant reading
NLP-based music knowledge discovery includes structured metadata extraction from web sources (with manually generated rules outperforming supervised learning for band-member detection) and sentiment analysis of 263,525 music reviews, which found more positive sentiment around 2008—attributed to circumstances including Obama's election—and a positive Reggae sentiment peak between 1975 and 1985 matching that genre's "golden age." Lyrics research ("lyrics information processing") spans rhyme-scheme identification, verse-chorus labeling, mood estimation, and topic modeling. Large-scale diachronic findings are consistent across studies: over 350,000 English song lyrics from 1970–2020 show declining lexical and structural complexity and increasingly negative emotional content; this corroborates earlier results on 6,150 Billboard Hot 100 songs (1951–2016) and more than 150,000 songs (1965–2010). A topic analysis of 1,364 Hip-Hop lyrics from the "rap wars" (1986–1998) distinguishes East coast ("Street Life and Rhythm") from West coast ("Vulgarity and Violence") topics. The paper also notes early LLM-based musicology studies reporting hallucination, opaque reasoning, and validity problems, though GPT-based evaluation of symbolic MIR tasks shows some promise.
Distant listening
Corpus studies dominate quantitative Pop musicology. Key results include:
- An analysis of 464,411 Western Pop songs (1955–2010) showing decreasing variety of pitch progressions, homogenizing timbre, and rising loudness (the "loudness war").
- A vocal-trend study of 145,912 tracks finding mean pitch increasing about one percent per year while pitch class entropy and total variation decline significantly—vocals are becoming less complex.
- A study of ~17,000 Billboard Hot 100 recordings using evolutionary-biology methods identifying three stylistic "revolutions" around 1964, 1983, and 1991; notably, its topic-level analysis contradicts the Serrà et al. finding, suggesting musical variety did not decrease but evolved through both continuous change and disruptions.
- A smaller (~1,000 recording) Top 40 study reporting increased minor-mode usage and decreased tempo, i.e., popular music sounding progressively sadder.
- Symbolic studies remain limited by data scarcity: the largest analyzed 1,131 MIDI melodies (top-5 Billboard songs, 1950–2023), finding three melodic revolutions consistent with declining complexity; automated transcription extended coverage to 1,571 chart songs, revealing slight increases in melodic repetition.
The paper explicitly flags methodological critiques here: several large studies rely on Million Song Dataset features that omit rhythm entirely and contain no audio, and the statistical detection of "revolutions" has been independently questioned.
Distant viewing
Computer vision work on cover art includes genre prediction from covers and promotional photos, genre-differentiated image features, and zero-shot object detection over 3,130 US chart album covers (1945–2003). Music video analysis remains niche; the largest resource contains only ~2,000 videos with features rather than raw media due to copyright. One study showed editors align shot transitions with downbeats for rhythmic genres like R&B and Reggaeton.
Multi-modal approaches
Genuinely multi-modal musicological work is rare. Notable exceptions include canonical correlation analysis over 119,664 lyric/audio pairs linking feature combinations to mood dimensions, a study of 124,288 metal lyrics correlating perceived audio hardness with lyrical brutality themes, and expert annotation of 21,000+ Billboard Hot 100 songs (1958–2022) with 58 attributes, yielding revolutions in 1964, 1983, and 2016 plus minor ones in 1991 and 2007—broadly consistent with Mauch et al. Audio LLMs (MuLan, LLARK, MusiLingo) are surveyed as emerging tools, with the open question of whether they hallucinate like text-only LLMs. Multi-modal embedding models trained jointly on audio, reviews, and cover art have shown, via class-activation heatmaps, that classifiers attend to faces for Rap, Blues, Reggae, R&B, Latin, and World genres but to instruments, typography, and clothing for Jazz.
Two open issues
Lack of multi-modality: despite the framing argument, only a small fraction of reviewed work combines modalities, and image/video modalities are severely under-represented relative to audio and lyrics.
Questionable external validity: corpora are built either from hit charts (limiting generalization to commercially successful, mostly US music) or from convenience samples dictated by availability rather than design. The paper stresses that large samples inflate statistical significance, so sampling biases can produce spurious significant results—a point grounded in broader critiques of big-data humanities and MIR-specific discussions of validity. It concedes that sampling biases cannot be fully eliminated and suggests industry collaboration (e.g., the Music Genome Project-derived dataset) as one partial remedy.
Three proposed research goals
First, charting the topic universe of Pop lyrics: scaling topic modeling (LDA or contextualized topic models) to millions of lyrics with temporal metadata, connecting topics to genre and societal trends—for example, testing whether female-fronted Metal bands exhibit lyrical topics differing from documented misogyny in the genre.
Second, an iconography of album cover art: combining automatic captioning (BLIP) and open-set object detection (Grounding DINO) to quantify motifs at scale, including sexualization indicators extending prior magazine-cover analyses, while acknowledging the cross-depiction problem—detectors trained on photographs struggle with drawn or painted cover art.
Third, tracking retro cycles: modeling cyclic genre revivals (e.g., Punk's debt to 1960s garage rock and its 1990s resurgence) via a joint multi-modal embedding space over audio, reviews, and images, ordered chronologically into distance matrices with Foote novelty formalization. A key claimed advantage is that joint modeling can reveal modality-specific influence—an artist may be visually close to Punk while musically distant—which parallel single-modality analyses cannot capture; ablation masking of individual modalities would quantify each modality's contribution.
Limitations and open questions
The paper is a position piece rather than an empirical study: its three research programs are sketched, not executed, and their feasibility at million-item scale is untested. The external-validity critique applies reflexively to the proposed agendas, since no representative large-scale Pop corpus currently exists. Whether Audio LLMs hallucinate at rates incompatible with musicological use remains unresolved, as does whether the conflicting findings on long-term variety trends reflect genuine methodological differences or corpus artifacts.
Conclusion
The article consolidates a computational digital humanities perspective on Pop music culture, documenting that substantial cultural engagement is exceptional within MIR's largely application-driven agenda. Its principal contributions are a systematic mapping of distant reading/listening/viewing results, an explicit diagnosis of multi-modality scarcity and sampling bias, and three concrete, technically specified research directions intended to reorient part of MIR's effort toward empirical cultural questions.
