Generic Symbolic Music Labeling Pipeline

Abstract

The availability of large datasets is an essential key factor for machine learning success. However, for symbolic music datasets, while there are many symbolic music files available, labeled datasets are scarce in many applications. In this paper, we propose a general pipeline for symbolic music labeling. The input to the pipeline is unlabeled midi files without particular constraints. Firstly, the pipeline filters the input and splits it into time-limited musical segments. Secondly, the pipeline generates intermediate labels using multiple pre-trained models, neural networks, and heuristics. Finally, multiple methods are used to combine intermediate labels to generate final labels. A label is accepted only if it exceeds a certain confidence level. To test the pipeline, we apply it to label a new piano difficulty dataset, "PianoDiff." We provide a thorough analysis to facilitate its usage in piano difficulty estimation for classification and generation using machine learning approaches. We test our pipeline on a dataset with manual labels. A random forest model trained on the weakly labeled dataset achieves an F1 score with a relative improvement of 13 percentage points compared to the same model trained on a smaller manually labeled dataset.

Emotion recognition from MIDI musical file using Enhanced Residual Gated Recurrent Unit architecture

“…Symbolic musical representation, similar to language modeling, involves converting MIDI files into discrete sequences of notes, mirroring musical events in a format akin to vocabulary (YGhatas et al, 2022). Tools like PrettyMIDI are used to extract specific details, such as each note's pitch, velocity, and duration.…”