DCUSV: deep clustering of ultrasonic vocalizations in rodents

Abstract Analyzing ultrasonic vocalizations (USVs) is critical for understanding rodents’ emotional states and social behaviors. This work presents Deep Clustering of USVs (DCUSV), an automated deep clustering pipeline for analyzing preprocessed USV contours that addresses key challenges in effectively clustering USVs and revealing distinct patterns in rodent vocal behavior. DCUSV employs a dense autoencoder to compress high-dimensional spectrograms into a latent space suitable for clustering, followed by a combination of Uniform Manifold Approximation and Projection (UMAP), Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN), and Agglomerative clustering, with hyperparameter optimization, to group USVs based on their spectro-temporal features. Clustering is evaluated using the Silhouette Coefficient, Calinski-Harabasz Index, and Davies-Bouldin Index. In addition, Principal Component Analysis (PCA), t-distributed Stochastic Neighbor Embedding (t-SNE), and UMAP are used to visualize and analyze the clustering results. Performance benchmarks against six baseline methods—K-Means, Deep Embedded Clustering (DEC), Improved Deep Embedded Clustering, Spectral Clustering, Gaussian Mixture Models, and HDBSCAN—revealed that DCUSV outperforms all baselines across every metric, achieving up to 2.62× higher Silhouette Coefficient scores, 22.95× higher Calinski-Harabasz scores, and 3.62× lower Davies-Bouldin scores (lower is better for the latter). Applying DCUSV revealed four distinct call-type families, closely aligning with manually defined categories without requiring manual grouping. Furthermore, DCUSV identified statistically significant shifts in call-type distributions across experimental conditions, demonstrating its ability to capture behaviorally meaningful changes in vocal expression. Thus, DCUSV enables robust analysis of USV structure and uncovers novel patterns in rodent vocal behavior.