Evaluating the Robustness of REMI-Tokenized and Regression-Based Symbolic Transformer under MIDI Distortions

The different touch methods on the piano can be identified as performance errors from distinct sound features, but it is still difficult to evaluate them solely through human intervention. Machine learning models can effectively solve this problem. This paper evaluates symbolic transformer models ability to automatically assess piano performance by utilizing controlled MIDI distortions. Using 1276 MAESTRO v3 excerpts, this research generates 27868 MIDI files categorized into eight distortion classes at three severity levels. This model is a REMI-tokenized regression-based transformer adaptation that is trained for 60 epochs on an 80/20 training and validation split. The regression model attains 63% validation accuracy with a validation loss of 0.28. The model demonstrates various levels of sensitivity to timing and pitch distortions. Results indicate that REMI tokenization and RegressLM integration improve detection of severe and moderate errors but struggle with subtle expressive deviations, highlighting directions for augmentation and curriculum learning to strengthen applicability.