Automatic Characterization of Flamenco Singing by Analyzing Audio Recordings

Flamenco singing is a highly expressive improvisational artform characterized by its deviation from the Western tonal system, freedom in rhythmic interpretation and a high amount of melodic ornamentation. Consequently, a singing performance represents a fusion of style-related constraints and the individual spontaneous interpretation. This study focuses on the description of the characteristics of a particular singer. In order to find suitable feature sets, a genre-specific automatic singer identification is implemented. For Western classical and popular music, related approaches have mainly relied on the extraction of timbre-based features to automatically recognize a singer by analyzing audio recordings. However, a classification solely based on spectral descriptors is prone to errors for low quality audio recordings. In order to obtain a more robust approach, low-level timbre features are combined with vibrato- and performance-related descriptors. Furthermore, differences among interpretations within a style are analyzed: Versions of the same a cappella cante have a common melodic skeleton which is subject to strong, individually determined melodic and rhythmic modifications. Similarity among such performances is modeled by applying dynamic time-warping to align automatic transcriptions and extracting performance-related descriptors. Resulting distances are evaluated by analyzing their correlation to human ratings.