Evaluation of Deep Learning Approaches for Pitch Scoring in Piano Practice and Performance

Abstract

This study evaluates deep learning approaches for pitch scoring in piano practice and performance through two experiments. The first experiment compares Gated Recurrent Units (GRU) and Transformer architectures using datasets that include diverse musical elements such as pitch, rhythm, rest, and tempo. The results demonstrate that Transformers significantly outperform GRUs in terms of accuracy and robustness across all conditions. The second experiment investigates modifications to the Transformer model, specifically increasing the number of attention heads, to assess its impact on transcribing musical sequences of varying complexity. Overall, these experiments high-light the strengths and limitations of Transformer architectures, emphasizing their potential to advance music transcription tools for education and professional applications.