---
language:
- en
license: cc-by-nc-4.0
tags:
- music
- audio
- multimodal
- regression
- aesthetics
- dataset
---

# MADB: Music Aesthetics Dataset and Benchmark

## Dataset Description

MADB is a large-scale dataset for music aesthetic evaluation, designed to support research on multi-dimensional and subjective music perception.

The dataset contains approximately **10,000 music tracks**, each annotated by **multiple trained annotators** across **10 perceptual dimensions** and one overall score. In addition, each track includes **textual comments** and **semantic tags** (genre and mood), enabling multimodal learning.

---

## Data Composition

### Audio

- All audio files are stored under `data/audio/` (mp3 format)  
- 1730 tracks are generated by Suno and Levo  
- 4400 tracks are from the Muchin dataset: https://github.com/CarlWangChina/MuChin  
- Remaining tracks are collected from diverse online sources  

---

### Annotations

Annotations are stored under `data/annotation/`:

- `avg_score.csv`: average score for each dimension per track  
- `MADB_data.csv`: full annotation data, including:
  - per-annotator scores  
  - textual comments  
  - genre and mood tags
- `split_val_seed5.csv`: the validate set with random seed 5

---

### Sample

Samples are stored under `sample/`:

- sample_audio: 200 audios selected from validate set:  
  - 100 audios randomly selected from Muchin   
  - 50 audios from levo  
  - 50 audios from suno
- sample_embedding: embedding extracted from sample audios:
  - clap: extracted with original clap  
  - muq: extracted with original muq
  - clap_com: extracted with clap after adaption with comments
  - clap_com_tag: extracted with clap after adaption with comments and tags
- `sample_ids.csv`: contains all sample audios' id

---

## Annotation Framework

Each track is rated across the following dimensions:

- Melody perception  
- Melody emotion  
- Arrangement perception  
- Arrangement emotion  
- Rhythm perception  
- Structure perception  
- Performance and singing mood  
- Enunciation and singing skill  
- Performance skill  
- Sound effect perception  
- Overall score  

Some dimensions may be **not applicable**, in which case a value of `0` is assigned.

---

## Annotation Process

- Annotators have at least **3 years of formal music training**  
- All annotators hold at least a **bachelor's degree**  
- Quality control is conducted by experts with **10+ years of professional experience**  
- Each track is rated by multiple annotators  

---

## Intended Use

This dataset is designed for:

- Music aesthetic evaluation  
- Multimodal learning (audio + text + tags)  
- Music understanding and analysis  
- Evaluation of generative music systems  


---

## Licensing and Usage

- Audio data may be subject to original copyright restrictions  
- Users should ensure compliance with the original data sources  
- This dataset is intended for **research purposes only**