MusPy: A Toolkit for Symbolic Music Generation (2008.01951v1)

Abstract: In this paper, we present MusPy, an open source Python library for symbolic music generation. MusPy provides easy-to-use tools for essential components in a music generation system, including dataset management, data I/O, data preprocessing and model evaluation. In order to showcase its potential, we present statistical analysis of the eleven datasets currently supported by MusPy. Moreover, we conduct a cross-dataset generalizability experiment by training an autoregressive model on each dataset and measuring held-out likelihood on the others---a process which is made easier by MusPy's dataset management system. The results provide a map of domain overlap between various commonly used datasets and show that some datasets contain more representative cross-genre samples than others. Along with the dataset analysis, these results might serve as a guide for choosing datasets in future research. Source code and documentation are available at https://github.com/salu133445/muspy .

• The paper introduces MusPy, an open-source Python toolkit that standardizes symbolic music generation through unified data management and conversion.
• It details a comprehensive approach integrating dataset management, interoperability with popular libraries, and support for various music representations.
• Experiments with autoregressive models highlight cross-dataset generalizability and refined evaluation metrics to enhance reproducible research.

MusPy: A Toolkit for Symbolic Music Generation

MusPy, presented in this paper, is an open-source Python library specifically designed for symbolic music generation. It addresses essential components in a music generation system, such as dataset management, data preprocessing, data input/output, and model evaluation. The development of MusPy emerges from the need for standardized tools that aid reproducibility and streamline the experimental process in music generation research.

Key Features

MusPy offers several features that distinguish it from other music manipulation libraries:

1. Data Management System: The toolkit includes a well-designed dataset management system inspired by the structure seen in TensorFlow Dataset and torchvision. This system covers eleven distinct datasets, offering an organized way to handle symbolic music data.
2. Unified Format with MusPy Music Class: A central element is the MusPy Music Class, a designed format that bridges existing file types like MIDI and MusicXML, enabling seamless conversions and manipulations. It retains the necessary data for music playback and symbolic notation.
3. Interoperability and Modularity: MusPy provides interface support for other popular symbolic music libraries, such as music21 and pretty_midi. This design choice ensures flexibility and modifiability within various machine learning frameworks.
4. Representation Support: MusPy supports different representations for symbolic music necessary for generative modeling, including pitch-based, event-based, note-based, and piano-roll representations.
5. Model Evaluation Tools: Tools for audio rendering, score visualization, and objective metrics assist researchers in assessing music generation systems. The objective metrics include pitch and rhythm-related measures that quantify the statistical properties of generated samples.

Dataset Analysis and Experimental Insights

MusPy's potential is demonstrated through an extensive analysis of the eleven datasets it currently supports. Analysis of song length, initial tempo, and key distributions highlights the variances and underscores the diversity among datasets. Additionally, several experiments using autoregressive models such as RNNs, LSTMs, GRUs, and Transformers elucidate complexities within individual datasets and test cross-dataset generalizability.

• General Observations: Models trained on larger, more diverse datasets like the Lakh MIDI Dataset (LMD) show broader generalizability compared to those trained on smaller, specific datasets.
• Cross-Dataset Generalizability: The experiments reveal variable generalizability between datasets, often influenced by the dataset's complexity and heterogeneity. Insights about model adaptability to various datasets provide guidance for training transfer models.
• Combining Datasets: An experimental approach that combines multiple datasets to create a unified dataset suggests improved model generalizability. Stratified sampling mitigates the imbalance bias caused by larger datasets like LMD, enhancing overall performance across various dataset types.

Implications and Future Directions

The implications of MusPy are manifold both practically and theoretically:

• Reproducibility: By providing standard tools and metrics, the toolkit improves reproducibility in symbolic music research, which is crucial for benchmark comparisons.
• Enhanced Research Efficiency: The modularity and flexibility of MusPy potentially reduce the time and effort required in the data preprocessing phase of music generation systems.
• Guidance for Dataset Selection: The dataset analysis and cross-dataset experiments provide actionable information for selecting suitable datasets in future studies.

Looking forward, MusPy could play a significant role in advancing AI-driven music composition by encouraging the creation and use of comprehensive, well-organized datasets. Future developments may focus on expanding dataset compatibility and refining model evaluation metrics to encompass more nuanced aspects of music, potentially leading to superior generative models and more sophisticated analyses.