Automatic Singer Identification For Improvisational Styles Based On Vibrato, Timbre And Statistical Performance Descriptors

Automatically detecting the singer by analyzing audio is a challenging task which gains in complexity for polyphonic material. Related approaches in the context of Western commercial music use machine learning models which mainly rely on low-level timbre descriptors. Such systems are prone to misclassifications when spectral distortions are present, since the timbre of the singer cannot be accurately modeled. For improvisational styles, where the performance is strongly determined by spontaneous interpretation characteristic for the singer, a more robust system can be achieved by additionally modeling the singer’s typical performance style. In addition to timbre and vibrato descriptors we therefore extract highlevel features related to the performance character from the predominant fundamental frequency envelope and automatic symbolic transcriptions. In a case study on flamenco singing, we observe an increase in accuracy for monophonic performances when classifying on this combined feature set. We furthermore compare the performance of the proposed approach for opera singing and investigate the influence of the album effect.