Symbolic-Domain Musical Instrument Classification Using Knowledge Distillation from Audio-Teacher to Symbolic-Student

This paper addresses the task of instrument classifi-cation focusing on musical note sequences such as sheet music or MIDI. Traditional research has primarily focused on instrument classification using audio signals, and the task of estimating performance instruments solely from note sequences has not been adequately addressed due to the difficulty of the task and the lack of labeled datasets. Typically, instrument classification models targeting audio signals learn and classify based on features such as timbre. This study aims to leverage the knowledge lost when converting audio signals to symbolic representations like musical scores, such as timbre information, for symbolic processing. We propose a method that performs knowledge distillation from a teacher model trained on audio signals to a student model trained on note sequences, enabling instrument classification with only note sequences as input. Experimental results demonstrate that utilizing the teacher's output distribution provides benefits beyond the regularization of label smoothing.