ccmusic-database/chest_falsetto

Dataset Card for Chest voice and Falsetto Dataset

The original dataset, sourced from the Chest Voice and Falsetto Dataset, includes 1,280 monophonic singing audio files in .wav format, performed, recorded, and annotated by students majoring in Vocal Music at the China Conservatory of Music. The chest voice is tagged as "chest" and the falsetto voice as "falsetto." Additionally, the dataset encompasses the Mel spectrogram, Mel frequency cepstral coefficient (MFCC), and spectral features of each audio segment, totaling 5,120 CSV files.

The original dataset did not distinguish between male and female voices, a critical detail for accurately identifying chest and falsetto vocal techniques. To correct this, we undertook a careful manual review and added gender annotations to the dataset. Following this process, we constructed the default subset of the current integrated version of the dataset, viewable in viewer. As the default subset had not undergone evaluation, we created the eval subset from it to verify the integrated dataset's effectiveness and completed the evaluation, viewable at [1]. Below is a brief overview of the data structure for each subset within the integrated dataset.

Dataset Structure

Default Subset Structure

audio	mel (spectrogram)	label (4-class)	gender (2-class)	singing_method (2-class)
.wav, 22050Hz	.jpg, 22050Hz	m_chest, m_falsetto, f_chest, f_falsetto	male, female	chest, falsetto

Eval Subset Structure

mel	cqt	chroma	label (4-class)	gender (2-class)	singing_method (2-class)
.jpg, 0.496s, 22050Hz	.jpg, 0.496s, 22050Hz	.jpg, 0.496s, 22050Hz	m_chest, m_falsetto, f_chest, f_falsetto	male, female	chest, falsetto

Data Instances

.zip(.wav, .jpg)

Data Fields

m_chest, f_chest, m_falsetto, f_falsetto

Data Splits

Split(6:2:2) / Subset	default & eval
train	767
validation	256
test	257
total	1280
total duration(s)	640.0513605442178

Usage

Default Subset

from datasets import load_dataset

ds = load_dataset("ccmusic-database/chest_falsetto")
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/chest_falsetto", 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_LFS_SKIP_SMUDGE=1 git clone [email protected]:datasets/ccmusic-database/chest_falsetto
cd chest_falsetto

Mirror

https://www.modelscope.cn/datasets/ccmusic-database/chest_falsetto

Dataset Description

Dataset Summary

For the pre-processed version, the audio clip was into 0.25 seconds and then transformed to Mel, CQT and Chroma spectrogram in .jpg format, resulting in 8,974 files. The chest/falsetto label for each file is given as one of the four classes: m chest, m falsetto, f chest, and f falsetto. The spectrogram, the chest/falsetto label and the gender label are combined into one data entry, with the first three columns representing the Mel, CQT and Chroma. The fourth and fifth columns are the chest/falsetto label and gender label, respectively. Additionally, the integrated dataset provides the function to shuffle and split the dataset into training, validation, and test sets in an 8:1:1 ratio. This dataset can be used for singing-related tasks such as singing gender classification or chest and falsetto voice classification.

Supported Tasks and Leaderboards

Audio classification, singing method classification, voice classification

Languages

Chinese, English

Dataset Creation

Curation Rationale

Lack of a dataset for Chest voice and Falsetto

Source Data

Initial Data Collection and Normalization

Zhaorui Liu, Monan Zhou

Who are the source language producers?

Students from CCMUSIC

Annotations

Annotation process

1280 monophonic singing audio (.wav format) of chest and falsetto voices, with chest voice tagged as chest and falsetto voice tagged as falsetto.

Who are the annotators?

Students from CCMUSIC

Considerations for Using the Data

Social Impact of Dataset

Promoting the development of AI in the music industry

Discussion of Biases

Only for chest and falsetto voices

Other Known Limitations

Recordings are cut into slices that are too short; The CQT spectrum column has the problem of spectrum leakage, but because the original audio slice is too short, only 0.5s, it cannot effectively avoid this problem.

Additional Information

Dataset Curators

Zijin Li

Evaluation

[1] https://huggingface.co/ccmusic-database/chest_falsetto

Contributions

Provide a dataset for distinguishing chest and falsetto voices