Detailseite
Projekt Druckansicht

Beiträge zur rechnergestützten Musikwissenschaft durch semi- und unüberwachtes Lernen zur Annotation und Segmentierung großer Musikarchive (ACMus)

Antragsteller Professor Dr.-Ing. Karlheinz Brandenburg, seit 12/2018
Fachliche Zuordnung Elektronische Halbleiter, Bauelemente und Schaltungen, Integrierte Systeme, Sensorik, Theoretische Elektrotechnik
Musikwissenschaften
Förderung Förderung von 2018 bis 2023
Projektkennung Deutsche Forschungsgemeinschaft (DFG) - Projektnummer 403542342
 
Erstellungsjahr 2022

Zusammenfassung der Projektergebnisse

The research project Advancing Computational Musicology: Semi-supervised and unsupervised segmentation and annotation of musical collections (ACMus) was an international and interdisciplinary collaboration between Colombian and German research institutes with a team of data scientist, professional musicians, and signal processing experts. The goal of ACMus was the creation of music analysis tools using semi- and unsupervised deep learning techniques for content with sparse available data. Hereby the main fields of interest were the classification of musical content based on the number of instruments, the rhythm, the vocals, the scale, and the meter. A Colombian archive of traditional music from the Andes region was the focus of research due to its challenging composition of unique rhythms and instrumentation, as well as the varying recording conditions. As a first step, an annotated dataset was created and published as the ACMUS-MIR dataset, and was continuously extended during the project. To ensure optimal annotations the process was led by professional musicians from Colombia. This annotations phase illustrated the expensive and time-consuming nature of this process even for small amounts of data, and emphasized the need of automated alternatives. Initial and promising results were achieved by applying the Transfer Learning technique to the classification tasks. This method trains neural networks in a data-rich domain to ensure that a sufficient amount of data is present for robust results. Subsequently, the knowledge (i.e. the trained network with its parameters) is transferred to the actual target task and fine-tuned on the sparse data. Furthermore, semi-supervised approaches initially developed for the field of computer vision, which include unlabeled data into the training process, were adapted to audio classification tasks. All of the applied methods significantly improved the classification performance and showed great potential for related future applications. With the ACMUS-MIR dataset a new annotated music collection is provided to the scientific community, bringing the focus to the unique regional and traditional music from the Andes. The applied techniques and classification models developed within the project show great potential for future research and applications. There are challenges for future work in simplifying the tools to make them usable by non-technical experts and allowing the adaption to novel domains. Future work can shift the focus to related areas of sparse data to evaluate the generalization of these methods and enlarge the potential user community.

Projektbezogene Publikationen (Auswahl)

 
 

Zusatzinformationen

Textvergrößerung und Kontrastanpassung