---
extra_gated_prompt: 'The song_structure dataset contains links to web audios used for data
  collection purposes. song_structure does not own or claim rights to the content linked
  within this dataset; all rights and copyright remain with the respective content
  creators and channel owners. Users are responsible for ensuring compliance with
  the terms and conditions of the platforms hosting these audios.'
extra_gated_fields:
  I acknowledge that song_structure does not own the audios linked in this dataset: checkbox
  I acknowledge that song_structure is not the original creator of the audios in this dataset: checkbox
  ? I understand that song_structure may modify or remove dataset content at the request
    of content creators or in accordance with platform policies
  : checkbox
  I accept the dataset license terms (CC-BY-NC-ND-4): checkbox
  I agree to use this dataset for non-commercial use ONLY: checkbox
license: cc-by-nc-nd-4.0
task_categories:
- time-series-forecasting
language:
- en
tags:
- music
- art
pretty_name: Song Structure Annotation Database
size_categories:
- n<1K
dataset_info:
  - config_name: default
    features:
      - name: audio
        dtype:
          audio:
            sampling_rate: 22050
      - name: mel
        dtype: image
      - name: label
        sequence:
          - name: onset_time
            dtype: uint32
          - name: offset_time
            dtype: uint32
          - name: structure
            dtype: string
    splits:
      - name: train
        num_bytes: 175969
        num_examples: 300
    download_size: 2308839939
    dataset_size: 175969
configs:
  - config_name: default
    data_files:
      - split: train
        path: default/train/data-*.arrow
---

# Dataset Card for Song Structure
The raw dataset comprises 300 pop songs in .mp3 format, sourced from the NetEase music, accompanied by a structure annotation file for each song in .txt format. The annotator for music structure is a professional musician and teacher from the China Conservatory of Music. For the statistics of the dataset, there are 208 Chinese songs, 87 English songs, three Korean songs and two Japanese songs. The song structures are labeled as follows: intro, re-intro, verse, chorus, pre-chorus, post-chorus, bridge, interlude and ending. Below figure shows the frequency of each segment label that appears in the set. The labels chorus and verse are the two most prevalent segment labels in the dataset and they are the most common segment in Western popular music. Among them, the number of “Postchorus” tags is the least, with only two present.

![](https://www.modelscope.cn/datasets/ccmusic-database/song_structure/resolve/master/data/msa.png)

## Dataset Structure
<style>
  .datastructure td {
    vertical-align: middle !important;
    text-align: center;
  }
  .datastructure th {
    text-align: center;
  }
</style>
<table class="datastructure">
    <tr>
        <th>audio</th>
        <th>mel</th>
        <th>label</th>
    </tr>
    <tr>
        <td>.mp3, 22050Hz</td>
        <td>.jpg, 22050Hz</td>
        <td>{onset_time:uint32,offset_time:uint32,structure:string}</td>
    </tr>
</table>

### Data Instances
.zip(.mp3), .txt

### Data Fields
```txt
intro, chorus, verse, pre-chorus, post-chorus, bridge, ending
```

### Data Split
train

## Dataset Description
### Dataset Summary
Unlike the above three datasets for classification, this one has not undergone pre-processing such as spectrogram transform. Thus we provide the original content only. The integrated version of the dataset is organized based on audio files, with each item structured into three columns: The first column contains the audio of the song in .mp3 format, sampled at 22,050 Hz. The second column consists of lists indicating the time points that mark the boundaries of different song sections, while the third column contains lists corresponding to the labels of the song structures listed in the second column. Strictly speaking, the first column represents the data, while the subsequent two columns represent the label.

### Supported Tasks and Leaderboards
time-series-forecasting

### Languages
Chinese, English

## Usage
```python
from datasets import load_dataset

dataset = load_dataset(
    "ccmusic-database/song_structure",
    name="default",
    split="train",
    token="Your_Access_Token",
)
for item in dataset:
    print(item)
```

## Maintenance
```bash
GIT_LFS_SKIP_SMUDGE=1 git clone git@hf.co:datasets/ccmusic-database/song_structure
cd song_structure
```

## Mirror
<https://www.modelscope.cn/datasets/ccmusic-database/song_structure>

## Dataset Creation
### Curation Rationale
Lack of a dataset for song structure

### Source Data
#### Initial Data Collection and Normalization
Zhaorui Liu, Monan Zhou

#### Who are the source language producers?
Students from CCMUSIC

### Annotations
#### Annotation process
Students from CCMUSIC collected 300 pop songs, as well as a structure annotation file for each song

#### Who are the annotators?
Students from CCMUSIC

### Personal and Sensitive Information
Due to copyright issues with the original music, only features of audio by frame are provided in the dataset

## Considerations for Using the Data
### Social Impact of Dataset
Promoting the development of the AI music industry

### Discussion of Biases
Only for mp3

### Other Known Limitations
Most are Chinese songs

## Additional Information
### Dataset Curators
Zijin Li

### Evaluation
<https://huggingface.co/ccmusic-database/song_structure>

### Contributions
Provide a dataset for song structure