---
license: cc-by-4.0
task_categories:
  - audio-classification
tags:
  - music
  - emotion
  - valence
  - arousal
  - continuous-annotation
  - music-emotion-recognition
pretty_name: MERP — Music Emotion Recognition with Profile Information
size_categories:
  - 1K<n<10K
---

# MERP — Music Emotion Recognition with Profile Information

[![Paper](https://img.shields.io/badge/Paper-Sensors%202023-blue)](https://doi.org/10.3390/s23010382)
[![GitHub](https://img.shields.io/badge/GitHub-dorienh/MERP-black?logo=github)](https://github.com/dorienh/MERP)

Music is capable of conveying many emotions. The level and type of emotion perceived by a listener, however, is highly subjective. In this study, we present the **Music Emotion Recognition with Profile information dataset (MERP)**. This database was collected through Amazon Mechanical Turk (MTurk) and features dynamical valence and arousal ratings of 54 selected full-length songs. The songs were selected from the Free Music Archive using an innovative method — a Triple Neural Network with the OpenSmile toolkit — to identify 50 songs with the most distinctive emotions, chosen to fully cover the four quadrants of the valence–arousal space. Four additional songs were selected from the DEAM dataset as anchor tracks for quality filtering. A total of 452 participants annotated the dataset, with 277 remaining after thorough cleaning. Their demographic information, listening preferences, and musical background were recorded.

Source code and training scripts: [github.com/dorienh/MERP](https://github.com/dorienh/MERP)

![MERP dataset](https://github.com/dorienh/MERP/blob/master/merp.png?raw=true "MERP Dataset")

---

## Corpus

| Property | Value |
| --- | --- |
| Tracks | 54 full-length songs |
| Track source | 50 tracks from the [Free Music Archive](https://freemusicarchive.org/) (CC-licensed); 4 anchor tracks from [DEAM](https://cvml.unige.ch/databases/DEAM/) (also FMA/CC) |
| Annotation resolution | 10 Hz (one rating per 0.1 s) |
| Annotation scale | [−1, 1] for both valence and arousal |
| MTurk batches | 4–6 (US, Master workers) and 7–8 (worldwide, non-Master) |
| Total raters (raw) | 447 |
| Total raters (filtered) | 277 |
| Raters with profile data | 196 |

Song IDs follow the pattern `{folder}_{track}` (e.g., `00_35`, `deam_115`). Audio files are in `audio/{song_id}.wav`.

---

## Files

| File | Rows | Description |
| --- | --- | --- |
| `annotations_filtered.parquet` | 3 482 | Quality-filtered annotations; per-rater rescaled to [−1, 1] and Savitzky-Golay smoothed (window 15, poly 2). This is the version used in the paper. |
| `annotations_raw.parquet` | 6 176 | Raw MTurk slider output before any filtering or rescaling. |
| `rater_profiles.parquet` | 196 | Demographic and musical background for raters who passed profile quality checks. |
| `audio/*.wav` | 54 | Full-length audio tracks. |

Rater IDs (`rater_001` … `rater_447`) are consistent across all three tables. Worker IDs from Amazon Mechanical Turk have been anonymized.

---

## Filtering steps (filtered split)

The `annotations_filtered` split had 7 quality-control steps applied:

1. Trials that were too short or too long relative to the audio duration were removed.
2. Trials from raters with erroneous profile entries (e.g., training duration < 0 or > 100 years) were removed.
3. Trials with long stagnant plateaus (≥ 300 identical consecutive values) were removed.
4. Raters who did not complete all 4 DEAM anchor songs were removed.
5. Raters whose valence or arousal range across all their trials was less than 1/3 of the full scale were removed.
6. Annotation arrays were trimmed to match the audio feature length.
7. Where a rater completed duplicate DEAM anchor sessions, the session with lower correlation to the DEAM reference annotations was discarded.

The original processing pipeline then applied **per-rater rescaling** (each rater's min → −1, max → 1) and **Savitzky-Golay smoothing** (window=15, polynomial order=2) to produce the values stored in this split.

---

## Column descriptions

### `annotations_filtered.parquet` / `annotations_raw.parquet`

| Column | Type | Description |
| --- | --- | --- |
| `rater_id` | str | Anonymized rater identifier (e.g., `rater_042`). Consistent with `rater_profiles.parquet`. |
| `song_id` | str | Track identifier matching the `audio/` file stem (e.g., `00_35`, `deam_115`). |
| `batch` | str | MTurk batch number (`"4"`–`"8"`). Batches 4–6 are US Master workers; 7–8 are worldwide non-Master. |
| `arousal` | list[float] | Time-series arousal ratings at 10 Hz. |
| `valence` | list[float] | Time-series valence ratings at 10 Hz. |

### `rater_profiles.parquet`

| Column | Type | Description |
| --- | --- | --- |
| `rater_id` | str | Anonymized rater identifier. |
| `master` | int | 1 = MTurk Master worker, 0 = non-Master. |
| `age` | int | Rater's age in years. |
| `country_enculturation` | str | Country where the rater grew up. |
| `country_live` | str | Country where the rater currently lives. |
| `fav_music_lang` | str | Preferred music language (e.g., `EN`, `ZH`). |
| `gender` | str | Self-reported gender. |
| `fav_genre` | str | Favourite music genre. |
| `play_instrument` | str | Whether the rater plays an instrument (`Yes`/`No`). |
| `training` | str | Whether the rater has formal music training (`Yes`/`No`). |
| `training_duration` | int | Years of formal music training (0 if none). |

---

## Usage

```python
import pandas as pd

# Load filtered annotations
ann = pd.read_parquet("annotations_filtered.parquet")

# Average valence per song across all raters
avg = ann.groupby("song_id")["valence"].apply(
    lambda arrs: [sum(x)/len(arrs) for x in zip(*arrs)]
)

# Join with rater profiles
profiles = pd.read_parquet("rater_profiles.parquet")
merged = ann.merge(profiles, on="rater_id", how="left")
```

---

## License

**Annotations and rater profiles:** [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)

**Audio files:** All tracks are from the Free Music Archive under Creative Commons licenses (CC BY, CC BY-SA, or CC BY-NC depending on the track). Please check the FMA metadata for per-track license details.

---

## Citation

```bibtex
@article{koh2023merp,
  title   = {{MERP}: A Music Dataset with Emotion Ratings and Raters' Profile Information},
  author  = {Koh, En Yan and Cheuk, Kin Wai and Heung, Kwan Yee and Agres, Kat R. and Herremans, Dorien},
  journal = {Sensors},
  volume  = {23},
  number  = {1},
  pages   = {382},
  year    = {2023},
  doi     = {10.3390/s23010382}
}
```