ccmusic-database/pianos

Dataset Card for Piano Sound Quality Dataset

The original dataset is sourced from the Piano Sound Quality Dataset, which includes 12 full-range audio files in .wav/.mp3/.m4a format representing seven models of pianos: Kawai upright piano, Kawai grand piano, Young Change upright piano, Hsinghai upright piano, Grand Theatre Steinway piano, Steinway grand piano, and Pearl River upright piano. Additionally, there are 1,320 split monophonic audio files in .wav/.mp3/.m4a format, bringing the total number of files to 1,332. The dataset also includes a score sheet in .xls format containing subjective evaluations of piano sound quality provided by 29 participants with musical backgrounds.

Based on the aforementioned original dataset, after data processing, we constructed the default subset of the current integrated version of the dataset, and its data structure can be viewed in the viewer. Due to the need to increase the dataset size and the absence of a popular piano brand, Yamaha, the default subset is expanded by recording an upright Yamaha piano into the 8_class subset. Since the current dataset has been validated by published articles, based on the 8_class subset, we adopted the data processing method for dataset evaluation from the article and constructed the eval subset, whose result has been shown in pianos. Except for the default subset, the rest of the subsets are not represented in our paper. Below is a brief introduction to the data structure of each subset.

Dataset Structure

Default Subset

audio	mel	label (8-class)	pitch (88-class)
.wav, 44100Hz	.jpg, 44100Hz	PearlRiver / YoungChang / Steinway-T / Hsinghai / Kawai / Steinway / Kawai-G / Yamaha	88 pitches on piano
...	...	...	...

Eval Subset

mel	label (8-class)	pitch (88-class)
.jpg, 0.18s 44100Hz	PearlRiver / YoungChang / Steinway-T / Hsinghai / Kawai / Steinway / Kawai-G / Yamaha	88 pitches on piano
...	...	...

Data Instances

.zip(.wav, jpg)

Data Fields

1_PearlRiver
2_YoungChang
3_Steinway-T
4_Hsinghai
5_Kawai
6_Steinway
7_Kawai-G
8_Yamaha (For Non-default subset)

Data Splits for Eval Subset

Split	Default	8_class	Eval
train(80%)	461	531	14678
validation(10%)	59	68	1835
test(10%)	60	69	1839
total	580	668	18352
Total duration(s)	2851.6933333333354	3247.941395833335	3247.941395833335

Viewer

https://www.modelscope.cn/datasets/ccmusic-database/pianos/dataPeview

Usage

Default Subset

from datasets import load_dataset

ds = load_dataset("ccmusic-database/pianos", name="default")
for item in ds["train"]:
    print(item)

for item in ds["validation"]:
    print(item)

for item in ds["test"]:
    print(item)

8-class Subset

from datasets import load_dataset

ds = load_dataset("ccmusic-database/pianos", name="8_classes")
for item in ds["train"]:
    print(item)

for item in ds["validation"]:
    print(item)

for item in ds["test"]:
    print(item)

Eval Subset

from datasets import load_dataset

ds = load_dataset("ccmusic-database/pianos", name="eval")
for item in ds["train"]:
    print(item)

for item in ds["validation"]:
    print(item)

for item in ds["test"]:
    print(item)

Maintenance

git clone git@hf.co:datasets/ccmusic-database/pianos
cd pianos

Dataset Summary

Due to the need to increase the dataset size and the absence of a popular piano brand, Yamaha, the dataset is expanded by recording an upright Yamaha piano in [1], in which the recording details can be found. This results in a total of 2,020 audio files. As models used in that article require a larger dataset, data augmentation was performed. The original audio was transformed into Mel spectrograms and sliced into 0.18-second segments, a parameter chosen based on empirical experience. This results in 18,745 spectrogram slices. Although 0.18 seconds may seem narrow, this duration is sufficient for the task at hand, as the classification of piano sound quality does not heavily rely on the temporal characteristics of the audio segments.

Supported Tasks and Leaderboards

Piano Sound Classification, pitch detection

Languages

English

Dataset Creation

Curation Rationale

Lack of a dataset for piano sound quality

Source Data

Initial Data Collection and Normalization

Zhaorui Liu, Shaohua Ji, Monan Zhou

Who are the source language producers?

Students from CCMUSIC & CCOM

Annotations

Annotation process

Students from CCMUSIC recorded different piano sounds and labeled them, and then a subjective survey of sound quality was conducted to score them.

Who are the annotators?

Students from CCMUSIC & CCOM

Personal and Sensitive Information

Piano brands

Considerations for Using the Data

Social Impact of Dataset

Help develop piano sound quality scoring apps

Discussion of Biases

Only for pianos

Other Known Limitations

Lack of black keys for Steinway, data imbalance

Additional Information

Dataset Curators

Zijin Li

Evaluation

[1] Monan Zhou, Shangda Wu, Shaohua Ji, Zijin Li, and Wei Li. A Holistic Evaluation of Piano Sound Quality[C]//Proceedings of the 10th Conference on Sound and Music Technology (CSMT). Springer, Singapore, 2023.

(Note: this paper only uses the first 7 piano classes in the dataset, its future work has finished the 8-class evaluation)

Citation Information

@dataset{zhaorui_liu_2021_5676893,
  author       = {Monan Zhou, Shenyang Xu, Zhaorui Liu, Zhaowen Wang, Feng Yu, Wei Li and Baoqiang Han},
  title        = {CCMusic: an Open and Diverse Database for Chinese and General Music Information Retrieval Research},
  month        = {mar},
  year         = {2024},
  publisher    = {HuggingFace},
  version      = {1.2},
  url          = {https://huggingface.co/ccmusic-database}
}

Contributions

Provide a dataset for piano sound quality