Music Source Separation on Classical Guitar Duets

Abstract

The dissertation presents an in-depth study on the separation of sources in classical guitar duets, addressing the unique challenge posed by the similar timbral characteristics of the instruments involved. This research introduces two new datasets comprised of real and synthetic recordings of guitar duets, designed to facilitate the exploration and evaluation of source separation techniques. Furthermore we propose a novel augmentation technique, OppositePanning, to enhance the separation process by exploiting the spatial distribution of sound, thereby offering a new avenue for improving source separation in settings where instruments share similar timbral characteristics. We propose a model pipeline which is is motivated by the understanding that guitar duet separation is inherently a hybrid task for humans. In practice, a human listener would naturally perceive the symbolic score from the audio, leveraging this score to aid in the separation process. This insight forms the foundation for the proposed dual-model pipeline, which aims to mimic this human approach by incorporating symbolic musical information directly into the separation algorithm. This approach is a significant departure from traditional source separation techniques, which primarily focus on the acoustic signal without considering the underlying musical structure. By employing a comparative analysis of Signal-to-Distortion Ratio (SDR) metrics, we evaluate the perfor- mance of the proposed dual-model pipeline against traditional methods. The findings demonstrate that incorporating symbolic musical information significantly improves separation accuracy, highlighting the im- portance of considering the musical context in source separation tasks. Moreover, this research posits that the methodologies and insights gained from the study of classical guitar duet separation could potentially be applied in other related fields as of speaker separation and voice singing separation, offering new perspectives and techniques for achieving more robust separation in complex auditory environments. The exploration of OppositePanning and the dual-model pipeline not only advances the understanding and methodology of monotimbral music source separation but also opens up avenues for further research in polyphonic music analysis and beyond, potentially leading to significant improvements in various applications of audio separation technology.