GiantMIDI-Piano: A large-scale MIDI dataset for classical piano music (2020-10-11T00:00:00.000000Z)

TL;DR

GiantMIDI-Piano is the largest classical piano MIDI dataset so far including the nationalities, the number and duration of works of composers, and the chroma, interval, trichord and tetrachord frequencies of six composers from different eras are shown to show that it can be used for musical analysis.

Abstract

Symbolic music datasets are important for music information retrieval and musical analysis. However, there is a lack of large-scale symbolic dataset for classical piano music. In this article, we create a GiantMIDI-Piano dataset containing 10,854 unique piano solo pieces composed by 2,786 composers. The dataset is collected as follows, we extract music piece names and composer names from the International Music Score Library Project (IMSLP). We search and download their corresponding audio recordings from the internet. We apply a convolutional neural network to detect piano solo pieces. Then, we transcribe those piano solo recordings to Musical Instrument Digital Interface (MIDI) files using our recently proposed high-resolution piano transcription system. Each transcribed MIDI file contains onset, offset, pitch and velocity attributes of piano notes, and onset and offset attributes of sustain pedals. GiantMIDI-Piano contains 34,504,873 transcribed notes, and contains metadata information of each music piece. To our knowledge, GiantMIDI-Piano is the largest classical piano MIDI dataset so far. We analyses the statistics of GiantMIDI-Piano including the nationalities, the number and duration of works of composers. We show the chroma, interval, trichord and tetrachord frequencies of six composers from different eras to show that GiantMIDI-Piano can be used for musical analysis. Our piano solo detection system achieves an accuracy of 89\%, and the piano note transcription achieves an onset F1 of 96.72\% evaluated on the MAESTRO dataset. GiantMIDI-Piano achieves an alignment error rate (ER) of 0.154 to the manually input MIDI files, comparing to MAESTRO with an alignment ER of 0.061 to the manually input MIDI files. We release the source code of acquiring the GiantMIDI-Piano dataset at this https URL.

References43 items

ASAP: a dataset of aligned scores and performances for piano transcription

High-Resolution Piano Transcription With Pedals by Regressing Onset and Offset Times

Polyphonic Piano Transcription Using Autoregressive Multi-State Note Model

PANNs: Large-Scale Pretrained Audio Neural Networks for Audio Pattern Recognition

Approachable Music Composition with Machine Learning at Scale

Adversarial Learning for Improved Onsets and Frames Music Transcription

Cross-task learning for audio tagging, sound event detection and spatial localization: DCASE 2019 baseline systems

Enabling Factorized Piano Music Modeling and Generation with the MAESTRO Dataset

Computational Models of Expressive Music Performance: A Comprehensive and Critical Review

Music Transformer: Generating Music with Long-Term Structure

Onsets and Frames: Dual-Objective Piano Transcription

A Tutorial on Deep Learning for Music Information Retrieval

MidiNet: A Convolutional Generative Adversarial Network for Symbolic-Domain Music Generation

Creating a Multitrack Classical Music Performance Dataset for Multimodal Music Analysis: Challenges, Insights, and Applications

Metrics for Polyphonic Sound Event Detection

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Optical music recognition: state-of-the-art and open issues

International Music Score Library Project

Deep Sparse Rectifier Neural Networks

Multiple Fundamental Frequency Estimation by Modeling Spectral Peaks and Non-Peak Regions

MAPS - A piano database for multipitch estimation and automatic transcription of music

Content-Based Music Information Retrieval: Current Directions and Future Challenges

The Challenge of Optical Music Recognition

The Art of Inaccuracy: Why Pianists' Errors Are Difficult to Hear

Multiple viewpoint systems for music prediction

The 'USI', or Universal Synthesizer Interface

SUPRA: Digitizing the Stanford University Piano Roll Archive

Performance Error Detection and Post-Processing for Fast and Accurate Symbolic Music Alignment

Computational Music Analysis

Learning-Based Methods for Comparing Sequences, with Applications to Audio-to-MIDI Alignment and Matching

Unfolding the potential of computational musicology

A New Music Database Describing Deviation Information of Performance Expressions