Musical Instrument Recognition Using Machine Learning Approaches

This work focuses on musical instrument recognition in miscellaneous audio sources, a fundamental issue of the machine auditory system. We contrast traditional machine learning techniques against deep learning strategies in a clinical dataset solved under a supervised multi-label, multi-output classification paradigm. To enhance the performance of the system, we adjust the analysis parameters for window size, hop size, and label threshold levels. Furthermore, we explore the benefits of auxiliary feature sets as alternative data representations, specifically: Auditory features are based on two categories: (1) four-band spectral decomposition of stereo sound data, left, right, mid, and side channels and (2) the harmonic and residual model of the audio signal. Two methods for fusing the models trained on these representations are suggested and we use the best of these methods to increase recognition accuracy. These enhancements lead to a vast improvement; with a 14.25 % increase M-F score for the macro average and a 24.17 % improvement in the exact match ratio overall compared to the baseline.