Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,91 @@
|
|
1 |
-
---
|
2 |
-
license: cc-by-nc-sa-4.0
|
3 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: cc-by-nc-sa-4.0
|
3 |
+
---
|
4 |
+
# SongEval π΅
|
5 |
+
**A Large-Scale Benchmark Dataset for Aesthetic Evaluation of Complete Songs**
|
6 |
+
|
7 |
+
[](https://huggingface.co/datasets/ASLP-lab/SongEval)
|
8 |
+
[](https://creativecommons.org/licenses/by-nc-sa/4.0/)
|
9 |
+
|
10 |
+
---
|
11 |
+
|
12 |
+
## π Overview
|
13 |
+
|
14 |
+
**SongEval** is the first open-source, large-scale benchmark dataset designed for **aesthetic evaluation of complete songs**. It provides over **2,399 songs** (~140 hours) annotated by **16 expert raters** across **five perceptual dimensions**. The dataset enables research in evaluating and improving music generation systems from a human aesthetic perspective.
|
15 |
+
|
16 |
+
|
17 |
+
<p align="center"> <img src="assets/intro.png" alt="SongEval" width="800"/> </p>
|
18 |
+
|
19 |
+
---
|
20 |
+
|
21 |
+
## π Features
|
22 |
+
|
23 |
+
- π§ **2,399 complete songs** (with vocals and accompaniment)
|
24 |
+
- β±οΈ **~140 hours** of high-quality audio
|
25 |
+
- π **English and Chinese** songs
|
26 |
+
- πΌ **9 mainstream genres**
|
27 |
+
- π **5 aesthetic dimensions**:
|
28 |
+
- Overall Coherence
|
29 |
+
- Memorability
|
30 |
+
- Naturalness of Vocal Breathing and Phrasing
|
31 |
+
- Clarity of Song Structure
|
32 |
+
- Overall Musicality
|
33 |
+
- π Ratings on a **5-point Likert scale** by **musically trained annotators**
|
34 |
+
- ποΈ Includes outputs from **five generation models** + a subset of real/bad-case samples
|
35 |
+
|
36 |
+
<div style="display: flex; justify-content: space-between;">
|
37 |
+
<img src="assets/score.png" alt="Image 1" style="width: 48%;" />
|
38 |
+
<img src="assets/distribution.png" alt="Image 2" style="width: 48%;" />
|
39 |
+
</div>
|
40 |
+
|
41 |
+
|
42 |
+
---
|
43 |
+
|
44 |
+
## π Dataset Structure
|
45 |
+
|
46 |
+
Each sample includes:
|
47 |
+
|
48 |
+
- `audio`: WAV audio of the full song
|
49 |
+
- `gender`: male or female
|
50 |
+
- `aesthetic_scores`: dict of five human-annotated scores (1β5)
|
51 |
+
|
52 |
+
---
|
53 |
+
|
54 |
+
## π Use Cases
|
55 |
+
|
56 |
+
- Benchmarking song generation models from an aesthetic viewpoint
|
57 |
+
- Training perceptual quality predictors for song
|
58 |
+
- Exploring alignment between objective metrics and human judgments
|
59 |
+
|
60 |
+
---
|
61 |
+
|
62 |
+
## π§ͺ Evaluation Toolkit
|
63 |
+
|
64 |
+
We provide an open-source evaluation toolkit trained on SongEval to help researchers evaluate new music generation outputs:
|
65 |
+
|
66 |
+
π GitHub: [https://github.com/ASLP-lab/SongEval](https://github.com/ASLP-lab/SongEval)
|
67 |
+
|
68 |
+
---
|
69 |
+
|
70 |
+
## π₯ Download
|
71 |
+
|
72 |
+
You can load the dataset directly using π€ Datasets:
|
73 |
+
|
74 |
+
```python
|
75 |
+
from datasets import load_dataset
|
76 |
+
|
77 |
+
dataset = load_dataset("ASLP-lab/SongEval")
|
78 |
+
|
79 |
+
```
|
80 |
+
|
81 |
+
## π Acknowledgement
|
82 |
+
|
83 |
+
We sincerely thank the **Shanghai Conservatory of Music** for their expert guidance on music theory, aesthetics, and annotation design.
|
84 |
+
<p align="center"> <img src="assets/sy_logo.jpg" alt="Shanghai Conservatory of Music Logo" width="300"/> </p>
|
85 |
+
|
86 |
+
---
|
87 |
+
|
88 |
+
## π¬ Citation
|
89 |
+
Coming soon!
|
90 |
+
|
91 |
+
|