Mixed Music Instrument Classification Based on Instantaneous Frequency Analysis and Time-Variant Spectrum Information
Kuan-Yu Chen, Jian–Jiun Ding, Yuan-Kang Lee · 2024
In recent years, the domain of Music Information Retrieval (MIR) has risen to prominence as a significant and focal point of investigation within the expansive realm of the music industry. MIR, in its multifaceted nature, encompasses diverse facets, delving into areas such as pitch recognition and the identification of timbral qualities. The primary focus of this particular study is directed towards the nuanced task of timbre identification, specifically pertaining to musical instruments. The process of Instrument Identification, as undertaken in this study, entails the application of the Harmonic-Temporal Clustering (HTC) technique. HTC serves as a methodological approach that dissects audio signals into constituent acoustic events through a meticulous time-frequency analysis. These acoustic events, thereby discerned, find representation in a spatial context defined by harmonic and temporal characteristics within a two-dimensional framework. The subsequent step involves the categorization of diverse instruments utilizing the HTC methodology. This categorization is subject to further scrutiny through analytical methods such as K-nearest-neighbor (KNN), Bayesian approaches, and the Non-Negative Matrix Factorization (NMF) algorithm, all contributing to the overarching goal of facilitating comprehensive classification. The overarching objective of this research is to establish meticulously refined HTC analysis models specifically tailored for brass and string instruments. This endeavor seeks to realize a level of precision in instrument recognition that is notably high. The system, as its primary operational framework, harnesses the prowess of Support Vector Machine (SVM) machine learning techniques, strategically employed to elevate the accuracy of the classification process. Furthermore, a judicious application of feature reduction methods is implemented with the express purpose of optimizing the aforementioned accuracy in the classification of musical instruments.