In an empirical study of music information retrieval, researchers Salomea Przedlacka and Piotr Bilski demonstrate that straightforward partition models frequently outperform intricate probabilistic architectures when organizing music tracks by intrinsic acoustic properties. Examining content-based clustering across twenty-five thousand audio tracks from the Free Music Archive, the study reveals that classical K-Means algorithms applied to carefully curated handcrafted features yield superior structural compactness and separation compared to high-parameter Gaussian mixture models. Crucially, the authors found that incorporating temporal segments derived from graph-based Laplacian segmentation consistently produces higher variation across cluster centroids than global whole-track descriptors. This confirms that isolating structurally coherent musical passages significantly sharpens algorithmic differentiation between audio profiles. Furthermore, the discovered clusters exhibited only weak alignment with traditional genre tags, establishing that signal-driven clustering uncovers acoustic relationships that diverge from subjective human categorization.
The broader field of music information retrieval encompasses critical tasks including recommendation engines, automated playlist generation, and catalog organization, all of which depend heavily on reliable similarity metrics. Traditionally, industrial recommendation platforms have relied upon collaborative filtering or listener engagement metadata, tracking historical playback patterns and user profiles rather than sonic content. While recent advances in deep learning employ self-supervised contrastive learning and deep neural embeddings to model music similarity, these latent representations often lack transparency, demand vast labeled corpuses, and resist straightforward mechanical interpretation. In response, Przedlacka and Bilski focused on handcrafted low-level and mid-level acoustic descriptors derived directly from audio waveforms. Handcrafted features permit a controlled, interpretable investigation of musical structure, enabling systems to uncover meaningful catalog organization solely through acoustic properties. This content-based framework provides a scalable foundation for recommendation systems that operate completely independently of user behavior.
To construct their experimental pipeline, the researchers utilized the medium subset of the Free Music Archive, evaluating twenty-five thousand audio tracks across diverse musical traditions. Rather than extracting acoustic measurements uniformly across entire audio files—an approach that blends verse, chorus, and bridge sections into diluted global averages—the authors introduced Laplacian segmentation. This graph-based spectral framework identifies internally consistent musical fragments by constructing a filtered binary recurrence matrix alongside local continuity matrices, balancing long-range structural repetitions with adjacent temporal coherence. By calculating a normalized graph Laplacian from the sequence-augmented affinity matrix, the algorithm projects audio frames into a low-dimensional eigenspace where structurally homogeneous passages become separable. The method segmented tracks into characteristic durations of twenty to forty seconds, evaluated segment consistency via Silhouette scores, and selected the most frequent segment in each track as its representative structural anchor.
Using the librosa software library, the research team extracted an initial set of one thousand sixty-four acoustic features from audio waveforms, excluding all accompanying textual metadata. The feature architecture spanned three fundamental musical dimensions: timbre, tonality, and rhythm. Timbral texture was captured using Mel-frequency cepstral coefficients, spectral roll-off, spectral centroid, spectral contrast, and root-mean-square energy. Tonality and harmonic structure were represented through waveform and spectrogram chromagrams, Chroma Energy Normalized statistics, and tonal centroid tonnetz descriptors. Rhythmic properties were parameterized through tempo estimates in beats per minute, onset beat time-series, and rhythmic value distributions. To convert these time-series into fixed-length mathematical vectors, the authors calculated statistical moments, including mean, variance, and skewness, across both full tracks and isolated segments. Before clustering, all extracted descriptors were independently scaled to a zero-to-one range using MinMax normalization fitted on the complete feature matrix, ensuring equal weighting across distance-based algorithms.
To eliminate redundant descriptors and mitigate the curse of dimensionality, the authors designed a multi-faceted feature selection strategy targeting approximately fifty features per method. Genre classification across ten balanced primary classes was adopted as an auxiliary supervised proxy to identify descriptors with high discriminative capability. The researchers applied minimum redundancy maximum relevance in mutual information difference mode, a heuristic Pearson correlation filter retaining features above a twelve percent correlation threshold, and statistical chi-square tests. Simultaneously, they trained a shallow LightGBM classifier with maximum tree depth of three and L1 and L2 regularization to derive gain-based feature importances, alongside TreeExplainer SHAP value estimations. Taking the union across all five selection mechanisms yielded a consolidated set of one hundred eighty-five unique features. This final subset preserved representation across all musical dimensions, integrating both global and segment-level descriptors such as segmented constant-Q chromagrams.
To verify that the chosen one hundred eighty-five descriptors retained meaningful musical information, the authors conducted a validation experiment using a multiclass LightGBM classifier. The model was trained in a one-versus-rest setup across ten random stratified holdout splits, utilizing a twenty percent test size with deterministic seeds incremented from forty-two. Across the ten independent splits, the model achieved an average area under the receiver operating characteristic curve of 0.97 on training data and 0.94 on testing data. This verified that the selected acoustic descriptors maintained strong discriminative power regarding musical characteristics. However, the researchers emphasized that because supervised genre labels guided feature screening, the resulting feature space carried an inherent inductive bias toward genre-correlated acoustic attributes. Consequently, the authors clarified that subsequent clustering results reflect organizational structures within a genre-informed feature space rather than a completely label-independent representation of acoustic content.
In the unsupervised clustering phase, the research investigated six algorithmic paradigms across cluster counts ranging from seven to forty. The tested models included partition-based K-Means and MiniBatch K-Means, hierarchical agglomerative clustering evaluating Ward, Complete, Average, and Single linkage, Gaussian Mixture Models, Birch, Fuzzy C-Means, and Self-Organizing Maps. To prevent clustering solutions from degenerating into severely skewed allocations dominated by a single massive cluster, the investigators introduced an empirical Gini coefficient threshold of 0.4. Because clustering algorithms do not inherently enforce uniform cluster sizes, this threshold effectively filtered out severely imbalanced outcomes. Under this balance criterion, Fuzzy C-Means and all agglomerative hierarchical methods except Ward linkage were excluded from further evaluation. This filtration demonstrated that flexible membership assignments and pairwise linkage heuristics frequently struggle to generate balanced partitions when operating over high-dimensional musical descriptor spaces.
To evaluate clustering quality without relying on ground-truth annotations, the researchers computed five internal validation metrics: the Silhouette score, Calinski-Harabasz index, Davies-Bouldin index, Dunn index, and normalized within-cluster sum-of-squares distortion. Although the traditional elbow method was examined to identify optimal cluster numbers, distortion curves decreased continuously across all tested values of k without displaying a clear inflection point, rendering the technique inconclusive. The authors therefore implemented a rank-based Borda aggregation procedure, independently ranking each algorithm configuration across all metrics and summing the ranks. K-Means and MiniBatch K-Means emerged as the most consistent performers across all evaluations, achieving their best aggregate ranks at k values of eleven and fifteen, respectively. Their high Silhouette and Calinski-Harabasz scores indicated that the underlying acoustic space aligns naturally with spherical centroid-based partitions, whereas Gaussian mixture models and Self-Organizing Maps showed substantially weaker cluster cohesion.
To directly evaluate feature relevance without training external surrogate classifiers, the study introduced an unsupervised centroid-variation analysis. For every clustering algorithm, the authors computed the standard deviation of each feature across cluster centroids, with higher variance indicating greater discriminative contribution to cluster separation. Crucially, across all tested algorithms, segmentation-derived features demonstrated consistently higher variation across centroids than global track descriptors. Gaussian Mixture Models at thirty-eight clusters exhibited the highest overall centroid variation and the strongest differentiation along segmented features, whereas hierarchical models like Birch and Ward linkage showed much lower variation. However, both K-Means and MiniBatch K-Means achieved substantially higher centroid variation than hierarchical methods despite utilizing far fewer clusters. This established that algorithmic architecture and model structure govern feature differentiation more decisively than cluster count alone, while confirming that temporal Laplacian segmentation provides structural information actively exploited by clustering models.
To evaluate whether the discovered clusters aligned with human musical concepts, the authors compared cluster memberships against original genre metadata using a normalized confusion matrix. With the partial exception of classical music, which mapped predominantly into a single cluster, genres exhibited broad dispersion across multiple clusters. Genres such as electronic, experimental, jazz, rock, pop, and international music spread extensively across cluster indices without forming isolated groupings. Similar weak alignment was observed across all tested clustering algorithms. These empirical observations confirm prior findings that musical genres rarely form acoustically cohesive clusters in unsupervised settings. Rather than reflecting subjective categorical labels, the clustering pipeline captured alternative structural affinities embedded in the acoustic signal. The authors emphasized that although the feature space was shaped using genre labels during feature selection, the unsupervised clustering process organized tracks according to acoustic similarity rather than reproducing predefined genre boundaries.
The investigation highlights a vital simplicity-versus-complexity trade-off in unsupervised music information retrieval. Although Gaussian Mixture Models possess greater representational capacity to parameterize multi-modal distributions, this flexibility did not yield superior clustering quality. In practice, K-Means at k equals eleven produced stronger internal validation metrics, whereas Gaussian Mixture Models at k equals thirty-eight generated high centroid variation that failed to translate into well-separated clusters due to substantial component overlap. However, because K-Means mathematically minimizes within-cluster squared distances, it inherently favors compactness metrics like Silhouette and Calinski-Harabasz, meaning that internal metrics alone cannot determine which model is perceptually superior. The authors identified key limitations, including a predefined k range, reliance on the Free Music Archive, and proxy genre bias during feature selection. Future work will include controlled ablation studies comparing performance with and without Laplacian segmentation, human listening evaluations, benchmark integration with Music4All-Onion, and self-supervised contrastive learning.
Subject of Research: Content-based clustering of musical audio using handcrafted features and temporal segmentation
Article Title: Content-based music clustering: evaluating features and comparative analysis
Article References: Przedlacka, S., & Bilski, P. (2026). Content-based music clustering: evaluating features and comparative analysis. Journal on Audio, Speech, and Music Processing, 2026(1), Article 40. https://doi.org/10.1186/s13636-026-00468-z
Image Credits: AI Generated
DOI: 10.1186/s13636-026-00468-z
Keywords: music information retrieval, content-based music clustering, audio feature extraction, k-means clustering, laplacian segmentation, free music archive, gaussian mixture models, feature selection, internal validation metrics, silhouette score, calinski-harabasz index, centroid variation analysis, spectral clustering, minimum redundancy maximum relevance, shap values, lightgbm, gini coefficient, acoustic similarity
Cite Scienmag News
Blake Davidson. (November 18, 2026). K-Means Outperforms Complex Audio Models When Guided by Segmented Musical Features. Scienmag. https://scienmag.com/k-means-outperforms-complex-audio-models-when-guided-by-segmented-musical-features/
Blake Davidson. "K-Means Outperforms Complex Audio Models When Guided by Segmented Musical Features." Scienmag, 18 November 2026, https://scienmag.com/k-means-outperforms-complex-audio-models-when-guided-by-segmented-musical-features/. Accessed 18 November 2026.
Blake Davidson. "K-Means Outperforms Complex Audio Models When Guided by Segmented Musical Features." Scienmag. November 18, 2026. https://scienmag.com/k-means-outperforms-complex-audio-models-when-guided-by-segmented-musical-features/








