Full-resolution PCA visualizations for feature representations in sound event classification and detection

This repository contains the full-resolution principal component analysis (PCA) visualizations generated for the study “Feature representations for sound event classification and detection in acoustic monitoring.” The associated study compares four audio feature representations for Sound Event Classification (SEC) and Sound Event Detection (SED): Contrastive Language-Audio Pretraining (CLAP) embeddings, mel-frequency cepstral coefficients (MFCC), gammatone-frequency cepstral coefficients (GFCC), and constant-Q cepstral coefficients (CQCC). The comparison is performed on two labeled environmental sound datasets, ESC-50 and DataSEC, using KMeans and Bisecting KMeans clustering. The repository includes one visualization for each combination of dataset, feature representation, and clustering algorithm. It therefore contains 16 full-resolution images: 2 datasets: ESC-50 and DataSEC; 4 feature representations: CLAP embeddings, MFCC, GFCC, and CQCC; 2 clustering algorithms: KMeans and Bisecting KMeans. Each image shows a two-dimensional PCA projection of the corresponding feature space. The plots include the reference sound-event classes, the clustering assignments obtained with the selected algorithm, cluster centroids, and convex hulls representing the spatial extent of the clusters in the projected space. The class identifiers reported in the plots correspond to the class coding adopted for the respective dataset. The visualizations are intended to support qualitative inspection of the spatial organization of the different feature representations. In particular, they allow the reader to examine cluster overlap, separation, compactness, centroid distribution, and the correspondence between the inferred clustering structure and the reference sound-event taxonomy. The plots complement the quantitative analysis reported in the associated manuscript, where clustering quality is evaluated using external, label-aware indices, namely the Adjusted Rand Index and the Fowlkes-Mallows index, and internal, label-agnostic indices, namely the Silhouette, Calinski-Harabasz, and Davies-Bouldin indices. The visualizations should not be interpreted as a complete representation of the original feature spaces. PCA reduces the data to two dimensions for graphical inspection and therefore introduces information loss. The clustering results and the conclusions of the study are based on the full analysis protocol described in the manuscript, while the images deposited here are provided as supporting qualitative material. The repository contains only the generated full-resolution visualizations. The original audio datasets are not redistributed and remain available through their respective public sources. File naming convention File names follow the structure: <dataset>_<feature>_<clustering_algorithm>.png where: <dataset> identifies the analyzed dataset, using ESC50 or dataSEC; <feature> identifies the feature representation: emb for CLAP embeddings; mfcc for mel-frequency cepstral coefficients; gfcc for gammatone-frequency cepstral coefficients; cqcc for constant-Q cepstral coefficients; <clustering_algorithm> identifies the clustering method, using kmeans or bisecting. Examples: dataSEC_cqcc_kmeans.pngdataSEC_cqcc_bisecting.pngdataSEC_emb_kmeans.pngESC50_mfcc_bisecting.png The code emb in the file names refers to the CLAP embedding representation analyzed in the associated manuscript. Intended use These files are provided to: enable full-resolution inspection of visualizations that cannot be reproduced clearly at journal page size; support comparison between handcrafted cepstral descriptors and CLAP embeddings; illustrate differences between KMeans and Bisecting KMeans partitions; complement the quantitative clustering validity indices reported in the manuscript; improve transparency and reproducibility of the qualitative component of the study. Related publication The visualizations are associated with the manuscript: L. Fredianelli, L. J. Pileggi, F. Artuso, and A. Akbaba, “Feature representations for sound event classification and detection in acoustic monitoring.”