Interpretable Feature Importance and Neural Embeddings for Classification and Similarity Analysis of One-Shot Percussive Sounds

The growing scale of one-shot percussive sample libraries makes manual organization impractical. This study investigates the relative contribution of hand-crafted spectral and temporal features to the classification of one-shot percussive sounds and examines how a learned embedding restructures the representation space. A 42-dimensional feature vector was extracted from 6,435 samples across seven instrument classes (kick, snare, clap, rims, hat, cymbal, bass), and a feed-forward neural network was trained with three configurations: the full 42 features, and the top-20 and top-10 subsets selected via gradient-based importance analysis. Across five independent runs, the importance rankings were highly stable (mean Spearman ρ = 0.949), with accumulated onset activity and energy variability ranking first and second in every run, ahead of a closely spaced group of cepstral and spectral-flux descriptors. Reducing the input from 42 to the top-20 features produced only a small decrease in validation accuracy (0.900→0.896); a further reduction to the top-10 features yielded a moderate drop (0.880). The learned embedding substantially improved cluster geometry (Silhouette: 0.058 → 0.498; Davies–Bouldin: 4.014 → 1.176) and class-consistent nearest-neighbor retrieval (Precision@5: 0.697→0.860). A training-fitted PCA baseline produced only marginal changes in cluster geometry and did not improve Precision@5, indicating that the observed ANN separation was not reproduced by linear variance-preserving projection alone. Three supplementary analyses delimit these results: no statistically significant difference was detected between the network and gradient boosting under the reported comparison protocol; no statistically significant sampling-rate improvement was detected for the residual Hat–Cymbal confusion; and external evaluation on 978 independently sourced samples showed smaller degradation for five classes but substantial degradation for Hat and Rims. On the reused validation partition, the top-20 configuration preserved most of the observed classification performance of the full feature set, while the learned embedding improved class-consistent sample organization. Because the same validation partition was used for early stopping, supervised feature ranking, and subset evaluation, the internal feature-subset comparisons should be interpreted as exploratory.