Violin perfomance analysis using weak supervision

This thesis investigates the development of core Music Information Retrieval (MIR) technologies for Violin Performance Analysis, addressing both data and label scarcity. To mitigate data scarcity, the Violin Repertoire Dataset, a large-scale collection of pedagogically motivated repertoire, is introduced. To address label scarcity, weak supervision techniques like regularized self training and forced alignment are utilized. The research focuses on developing pitch estimation, transcription, and playing technique detection algorithms for practical applications in music education. Additionally, the analysis considers direct estimation of violin performance parameters under piano accompaniment, a common scenario in exams and auditions. Quantitative results suggest that domain-specific applications, supported by the Violin Repertoire Dataset and weakly-supervised learning methods, deliver high-quality performances in several violin performance analysis tasks.