GiantMIDI-Piano: A large-scale MIDI dataset for classical piano music (2010.07061v3)

Abstract: Symbolic music datasets are important for music information retrieval and musical analysis. However, there is a lack of large-scale symbolic datasets for classical piano music. In this article, we create a GiantMIDI-Piano (GP) dataset containing 38,700,838 transcribed notes and 10,855 unique solo piano works composed by 2,786 composers. We extract the names of music works and the names of composers from the International Music Score Library Project (IMSLP). We search and download their corresponding audio recordings from the internet. We further create a curated subset containing 7,236 works composed by 1,787 composers by constraining the titles of downloaded audio recordings containing the surnames of composers. We apply a convolutional neural network to detect solo piano works. Then, we transcribe those solo piano recordings into Musical Instrument Digital Interface (MIDI) files using a high-resolution piano transcription system. Each transcribed MIDI file contains the onset, offset, pitch, and velocity attributes of piano notes and pedals. GiantMIDI-Piano includes 90% live performance MIDI files and 10\% sequence input MIDI files. We analyse the statistics of GiantMIDI-Piano and show pitch class, interval, trichord, and tetrachord frequencies of six composers from different eras to show that GiantMIDI-Piano can be used for musical analysis. We evaluate the quality of GiantMIDI-Piano in terms of solo piano detection F1 scores, metadata accuracy, and transcription error rates. We release the source code for acquiring the GiantMIDI-Piano dataset at https://github.com/bytedance/GiantMIDI-Piano

• The paper introduces a comprehensive dataset featuring over 38 million MIDI-transcribed notes from classical piano works.
• It employs advanced neural networks to transcribe both live and sequenced performances with high fidelity, achieving an F1 score of 88.14%.
• The dataset enables detailed analysis of composer-specific traits and historical trends, advancing research in music generation and computational musicology.

Towards a Comprehensive Resource for Symbolic Music: The GiantMIDI-Piano Dataset

The paper "GiantMIDI-Piano: A Large-Scale MIDI Dataset for Classical Piano Music" addresses the growing need for expansive symbolic datasets within the field of music information retrieval (MIR) and computational music analysis. While current offerings are limited in scope, the introduction of the GiantMIDI-Piano dataset aims to fill this gap by providing a vast collection of classical piano music MIDI files.

Dataset Creation and Composition

The essence of GiantMIDI-Piano lies in its scale and diversity. The dataset encompasses 38,700,838 transcribed notes from 10,855 unique solo piano works, attributed to 2,786 composers. By leveraging resources such as the International Music Score Library Project (IMSLP) and YouTube, audio recordings were identified and transcribed using advanced neural networks into MIDI files. The emphasis on high-transcription fidelity is underscored by the dataset's composition—90% live performances and 10% sequenced inputs.

A curated subset of the dataset further refines this collection by ensuring composer surname matches in the titles of recordings, resulting in 7,236 works by 1,787 composers. This detailed curation enhances the dataset's utility and reliability for diverse analytical purposes.

Analytical Potential and Evaluation

The authors provide a thorough statistical analysis of the dataset, revealing composer-specific traits and historical trends. For instance, the distribution of note pitches reflects traditional and modern uses of the piano range, highlighted by comparisons among composers like J.S. Bach and Debussy. This kind of insight can significantly enable both computational musicology and algorithm training for music generation.

The dataset's quality is rigorously evaluated through various metrics. Solo piano detection yields an F1 score of 88.14%, and transcriptions demonstrate a relative error rate of 0.094 against established benchmarks like the MAESTRO dataset.

Practical and Theoretical Implications

The introduction of GiantMIDI-Piano bears implications beyond straightforward music retrieval. Its extensiveness serves as a foundation for advancements in music generation, performance analysis, and symbolic music transcription. Furthermore, the ability to analyze extensive and varied datasets opens avenues for theoretical exploration into compositional styles, performance dynamics, and cultural evolution of musical works.

Future Directions

Despite its comprehensive nature, GiantMIDI-Piano acknowledges areas for potential improvement. These include integration of additional musical elements like beat, key, and interpretative nuances. As such, the dataset's growth will likely accompany further technological and methodological advances in MIR and AI music generation.

In conclusion, the GiantMIDI-Piano dataset represents significant progress in symbolic music research, offering a robust resource for both applied and theoretical pursuits. It stands poised to spur further innovation in the intersection of AI and musicology.