Boost Predominant Instrument Recognition Performance with MagiaSearch and MagiaClassifier

Hao Zhou,Zhen Li,Shusong Xing,Zujun Gu,Binhui Wang
DOI: https://doi.org/10.1007/978-3-031-44198-1_11
2023-01-01
Abstract:The objective of this study is to overcome the performance limitations of existing instrument recognition systems in a cost-effective manner. Identifying predominant instruments accurately is a critical problem in music information retrieval, and it directly affects the performance of various advanced techniques. To address this, we propose a novel instrument recognition system that integrates a fast search technique, named MagiaSearch, to discover reliable SpecAugment parameters applicable to instrument recognition and a deep net classifier, named MagiaClassifier, which uses Swin Transformer V2 as the backbone model. Our experiments demonstrate that MagiaSearch effectively searches for reliable SpecAugment parameters applied to log mel spectrograms of instrument audio, MagiaClassifier enhances the performance of instrument recognition systems, and combining MagiaSearch and MagiaClassifier, we achieve a significant accuracy of 88.76% for major instrument recognition tasks in 11 categories in the IRMAS dataset.
What problem does this paper attempt to address?