Interpreting Gender-Related Vocal Diversity through wav2vec 2.0 Speech Embedding Analysis

Self-supervised speech models such as wav2vec 2.0 produce high-dimensional embeddings that encode multiple aspects of vocal production, although the internal organization of these representations remains difficult to interpret. This work investigates the geometric structure of speaker-level speech embeddings extracted from a fine-tuned wav2vec 2.0 model using dimensionality reduction and acoustic analysis. Speaker representations were obtained from the Mozilla Common Voice 14.0 dataset and restricted to speakers in the \textit{twenties} and \textit{thirties} age groups in order to reduce age-related variability in vocal production. Embeddings were projected using Principal Component Analysis (PCA) and Uniform Manifold Approximation and Projection (UMAP), and the resulting spaces were analyzed visually and quantitatively through the Davies--Bouldin index. Additionally, the projections were compared with the median fundamental frequency F0 of each speaker to investigate the relationship between latent organization and interpretable acoustic information. The results show that both PCA and UMAP preserve meaningful acoustic structure, revealing gradual F0 transitions across the latent space. PCA produced more stable global organization, while UMAP emphasized local neighborhood structures. Cluster separability improved in binary classification settings, particularly in the \textit{male} versus \textit{female} configuration, while the three-class scenario exhibited stronger overlap. The analysis also indicated that the \textit{other} category appeared consistently closer to the \textit{female} class than to the \textit{male} class, suggesting asymmetric acoustic relationships between vocal categories. These findings contribute to a more interpretable understanding of self-supervised speech representations through the combined analysis of dimensionality reduction, acoustic descriptors, and latent space geometry.