Classification of piano performers with deep learning models

Abstract

Artists’ classification based on their unique performing style remains a rarely explored problem in music information retrieval. Nevertheless, it is an extremely intriguing subject that challenges even experienced experts. In this study, deep learning approaches, namely Convolutional Recurrent Neural Network (CRNN) and Transformer-based model, are employed to classify piano performers based on audio recordings. CRNN combines convolutional layers for the extraction of spectral characteristics with recurrent layers for temporal sequence modeling whereas Transformer-based models are widely used for sequence analysis, which makes both architectures suitable for music processing. To evaluate the ability of the model to recognize performers, multiple tests are conducted, with a varying number of both performers and composers of performed music, assessing how these variations impact accuracy. Depending on these aspects, the models’ accuracy ranges from $\sim 70\text{\%}$ to $\sim 98\text{\%}$. The findings show that pianist classification using deep learning is feasible, but could still be improved with further refinements. The potential applications include digital music archiving, historical performance analysis, or AI-assisted stylistic studies.