Abstract
In Rebetiko music, the distinctive vocal qualities of individual singers, especially renowned singers, are often at the centre of attention, as their unique timbre and ornamentations play a crucial role in defining the genre’s characteristics. Pre-trained deep learning models on human voice data offer a good starting point for fine-tuning to classify small-sized voice-related datasets. While these fine-tuned models are "black boxes" that do not provide direct explanations for their classifications, their internal latent representations can be explored through visualizations. This study analyzes the performance and possibilities for explainability of pre-trained voice models fine-tuned for identifying Rebetiko singers using separated voice signals from original recordings. The selected singers exhibit diverse timbral and delivery styles, which can be numerically represented through perceptually relevant audio features (e.g., fundamental frequency, spectral centroid, vibrato rate and depth, inharmonicity, etc.). By comparing extracted audio features with visualizations from the "black box" model, valuable insights into the model’s learning process are gained. This study not only demonstrates the effectiveness of fine-tuning for Rebetiko singer identification but also enhances explainability, which is crucial for better understanding the characteristics and trends of Rebetiko singing.
Funding Information
  • European Regional Development Fund of the European Union and Greek national funds (MIS 5149205)