Percussionists use a multitude of objects and materials, mounted on their instruments, to achieve a satisfying sound texture. This is a tedious process as there are no guidelines suggesting how to manipulate a percussion instrument to adjust its perceptual characteristics in the desired direction. To this end, the article presents a methodology for computationally identifying how to damp and tune a drumhead by adjusting its mass distribution, e.g., by applying malleable paste on its surface. A dataset of 11,114 sounds has been synthesized using a FDTD solution of the wave equation representing the vibration of a membrane, which is being transmuted through the application of paste. These sounds are investigated to derive conclusions concerning their spectral characteristics and data reduction techniques are used to investigate the feasibility of computationally inferring damping parameters for a given sound. Furthermore, these sounds are used to train a Convolutional Neural Network to infer mass distribution from sound. Results show that computational approaches can provide valuable information to percussionists striving to adjust their personal sound. Although this study has been performed with synthesized sounds, the research methodology presents some inspiring ideas for future investigations with prerecorded sounds.
New perspectives in correlating physical and perceptual characteristics of musical instruments have emerged through recent advancements in machine learning and high-performance computing. Driving such correlating processes, for instance through training deep learning models, requires the assembly of large audio datasets and the development of tools for Exploratory Data Analysis (EDA). This article presents AudioInsight, a web application for the audio-visual exploration of sound datasets for musical acoustics. This application is a by-product of our research on drumhead acoustics, which aims at the computational inference of damping strategies resulting in a desired sound texture. Developed in Python and Dash-Plotly, AudioInsight features an interactive and responsive user interface providing tools for dataset exploration, statistical analysis and visualization of parameter distributions and relevant correlations. Moreover, AudioInsight permits generating dataset clusters and exploring sounds by interactively navigating within their 2D or 3D graphic representations. Future enhancements of this application, allowing musicians and researchers to contribute to musical instrument datasets through crowdsourcing, may serve as an example application within the Internet of Sounds (IoS) ecosystem.
This study presents a representation-centric evaluation of audio foundation models for fine-grained musical instrument analysis, focusing on cymbal classification. A confound-aware comparison of CLAP and MERT embeddings is conducted to examine how each latent space supports recoverability of acoustically and semantically relevant information. To support this analysis, the study introduces a representation-centric, confound-aware multi-stage evaluation framework that separates exploratory geometry, leakage-safe probing, and supporting unsupervised clustering evidence. The methodology is applied to a challenging cymbal dataset characterized by hierarchical labels, class imbalance, and subtle acoustic variation. Results reveal a target-dependent profile of representational strengths rather than a single overall winner. CLAP exhibits stronger variance concentration and more label-consistent local neighborhood organization, and it outperforms MERT on fine-grained, strike-related targets. MERT, however, retains a small but consistent advantage on higher-level cymbal-type classification. Unsupervised analyses show that these advantages reflect local neighborhood structure, not strong global cluster formation, and confound diagnostics indicate that size-related information remains largely type-mediated. Overall, the findings underscore the importance of structured, multi-stage evaluation for disentangling embedding geometry, recoverability, and confound effects while demonstrating the complementary strengths of AFMs in complex audio classification settings.
Musical instrument analysis, focusing on cymbal classification. A confound-aware comparison of CLAP and MERT embeddings is conducted to examine how each latent space supports recoverability of acoustically and semantically relevant information. To support this analysis, a five-stage evaluation framework is introduced, comprising geometric inspection, leakage-safe probing, and unsupervised clustering, complemented by confound diagnostics formulated as a horizontal task across all stages. The methodology is applied to a challenging cymbal dataset characterized by hierarchical labels, class imbalance, and subtle acoustic variation. Results reveal a target-dependent profile of representational strengths rather than a single overall winner. CLAP exhibits stronger variance concentration and more label-consistent local neighborhood organization, and it clearly outperforms MERT on fine-grained, strike-related targets. MERT, however, retains a small but consistent advantage on higher-level cymbal-type classification. Unsupervised analyses show that these advantages reflect local neighborhood structure rather than strong global cluster formation, and confound diagnostics indicate that size-related information remains largely type-mediated. Overall, the findings underscore the importance of structured, multi-stage evaluation for disentangling embedding geometry, recoverability and confound effects, while demonstrating the complementary strengths of AFMs in complex audio classification settings.