GiantMIDI-Piano: A large-scale MIDI dataset for classical piano music
GiantMIDI-Piano dataset uses convolutional neural networks and high-resolution transcription, containing 38,700,838 notes.
Key Findings
Methodology
The study employs convolutional neural networks to detect solo piano works and transcribes audio into MIDI files using a high-resolution piano transcription system. The dataset includes 38,700,838 notes and 10,855 unique works. Composer and work names are extracted from IMSLP, and corresponding audio recordings are downloaded from the internet. A modified Jaccard similarity is used to filter audio, ensuring match accuracy.
Key Results
- Result 1: The GiantMIDI-Piano dataset contains 38,700,838 notes, covering 10,855 solo piano works by 2,786 composers, demonstrating its potential for music analysis.
- Result 2: The solo piano detection using convolutional neural networks achieved high F1 scores, with high metadata accuracy and low transcription error rates.
- Result 3: Analysis of pitch class, interval, trichord, and tetrachord frequencies of six composers from different eras validates the dataset's application in music analysis.
Significance
GiantMIDI-Piano fills the gap of large-scale symbolic datasets for classical piano music, providing a rich resource for music information retrieval and analysis. Its 90% live performance MIDI files and 10% sequence input MIDI files lay the foundation for musicological research and computer music generation.
Technical Contribution
This study significantly improves the detection and transcription accuracy of solo piano works through convolutional neural networks and high-resolution transcription systems. Compared to existing datasets, GiantMIDI-Piano offers significant enhancements in scale and diversity, supporting broader music analysis and generation tasks.
Novelty
GiantMIDI-Piano is the first large-scale classical piano MIDI dataset, combining automated audio downloading with high-precision transcription technology, offering unprecedented opportunities for music analysis.
Limitations
- Limitation 1: The dataset relies on internet audio, where quality and recording conditions may affect transcription accuracy.
- Limitation 2: Despite using high-resolution transcription systems, challenges remain in transcribing complex harmonies and fast note sequences.
- Limitation 3: The dataset's composer nationality information is incomplete, potentially affecting cross-cultural music analysis.
Future Work
Future research could expand to datasets for other instruments and music styles, further improve transcription system accuracy, and explore applications in music generation and automatic composition.
AI Executive Summary
The GiantMIDI-Piano dataset, utilizing convolutional neural networks and high-resolution transcription systems, provides a large-scale symbolic dataset for classical piano music. Existing datasets are often limited in scale and composer coverage, failing to meet the diverse needs of music analysis. GiantMIDI-Piano extracts composer and work information from IMSLP, downloads audio from the internet, detects solo piano works using convolutional neural networks, and transcribes them into MIDI files. Experimental results show excellent performance in solo piano detection and transcription accuracy, supporting applications in music information retrieval and generation. Despite limitations in audio quality and transcription complexity, the dataset offers a valuable resource for academic research and industrial applications, with potential expansion to other music styles and instruments in the future.
Deep Analysis
Background
The evolution of music information retrieval and analysis requires large-scale symbolic datasets. Existing piano MIDI datasets like piano-midi.de and MAESTRO are limited in scale, failing to cover a wide range of classical piano works. GiantMIDI-Piano provides a rich resource for music analysis by combining automated audio downloading and high-precision transcription technology.
Core Problem
The lack of large-scale symbolic datasets for classical piano music limits research in music information retrieval and analysis. Existing datasets are small in scale, with incomplete composer and work coverage, failing to meet diverse research needs.
Innovation
GiantMIDI-Piano constructs a large-scale classical piano MIDI dataset through automated audio downloading and high-precision transcription technology. Its innovation lies in using convolutional neural networks for solo piano detection, ensuring dataset accuracy and diversity.
Methodology
- �� Extract composer and work information from IMSLP
- �� Search audio on YouTube using keywords
- �� Use convolutional neural networks to detect solo piano works
- �� Transcribe audio into MIDI files using a high-resolution transcription system
- �� Analyze pitch class, interval, and chord frequencies
Experiments
The experimental design includes using convolutional neural networks to detect solo piano works and transcribing them into MIDI files using a high-resolution transcription system. The dataset quality is evaluated using F1 scores, metadata accuracy, and transcription error rates.
Results
The GiantMIDI-Piano dataset shows excellent performance in solo piano detection and transcription accuracy, with high F1 scores, metadata accuracy, and low transcription error rates. Analysis of six composers' pitch class, interval, and chord frequencies validates the dataset's application in music analysis.
Applications
GiantMIDI-Piano can be used in music information retrieval, music generation, music analysis, and automatic composition. Its rich coverage of composers and works provides a foundation for cross-cultural music research.
Limitations & Outlook
The dataset relies on internet audio, where quality and recording conditions may affect transcription accuracy. Challenges remain in transcribing complex harmonies and fast note sequences. The dataset's composer nationality information is incomplete, potentially affecting cross-cultural music analysis.
Plain Language Accessible to non-experts
Imagine you're in a giant music library with thousands of scores and recordings. GiantMIDI-Piano is like a super-smart librarian that can quickly find the scores you want and convert them into music files you can play on your computer. By using a technology called convolutional neural networks, it identifies which recordings are solo piano, then uses a high-resolution transcription system to turn these recordings into MIDI files, just like scanning books into e-books. This way, both music scholars and enthusiasts can easily analyze and study these musical works.
ELI14 Explained like you're 14
Imagine you have a super cool music robot that understands all the piano music in the world! GiantMIDI-Piano is like that robot. It can find all sorts of piano music online and use a technology called convolutional neural networks to figure out if they're solo piano pieces. Then, it uses a high-tech system to turn these pieces into MIDI files, just like turning your favorite game music into a format you can play on your computer. This way, you can do all sorts of fun things with the music, like creating your own tunes or studying different composers' styles.
Glossary
Convolutional Neural Network
A deep learning algorithm used for image and audio recognition, effectively extracting features.
Used to detect solo piano works.
MIDI (Musical Instrument Digital Interface)
A digital music file format containing pitch, duration, and velocity information of notes.
Transcription system converts audio to MIDI files.
IMSLP (International Music Score Library Project)
An online music library offering a large collection of public domain music works.
Used to extract composer and work information.
Transcription System
A system that converts audio signals into symbolic representations.
Used to transcribe piano recordings into MIDI files.
Jaccard Similarity
A metric for measuring the similarity of sets, calculating the ratio of intersection to union.
Used to filter audio match accuracy.
Open Questions Unanswered questions from this research
- 1 How to improve transcription accuracy for complex harmonies and fast note sequences? Current systems perform poorly in these areas.
- 2 How to expand the dataset to cover more music styles and instruments? The current dataset focuses mainly on classical piano.
- 3 How to ensure that internet audio quality and recording conditions do not affect transcription results?
Applications
Immediate Applications
Music Analysis
Researchers can use the dataset to analyze composer styles and music structures, supporting academic music research.
Music Generation
Music creators can use the dataset to train generative models and create new music compositions.
Long-term Vision
Automatic Composition
By expanding the dataset and improving transcription accuracy, develop automatic composition systems to support music education and creation.
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