README.md

---
license: cc-by-nc-4.0
language:
- en
tags:
- music
- audio-features
- spotify
- avd
- music-psychology
- emotion
- arousal-valence
pretty_name: Spotify AVD (Arousal · Valence · Depth) Aggregates
size_categories:
- 1M<n<10M
configs:
- config_name: artists
  data_files: artists_avd.parquet
- config_name: genres
  data_files: genres_avd.parquet
---

# Spotify AVD — Arousal · Valence · Depth

A large-scale, psychology-backed dataset of **artist- and genre-level musical attribute profiles**, computed from Spotify audio features. It pairs the well-known **Arousal–Valence** model with a derived **Depth** dimension, the three-attribute framework introduced by Greenberg et al. (2016) and later replicated with computer-extracted audio features by Fricke et al. (2018).

- **2,204,210 artists** with popularity-weighted AVD means, per-axis spread, follower counts, and track coverage.
- **734 genres** with the same aggregate statistics.
- Built from the audio-feature layer of the Spotify catalogue (~255M tracks) — this release publishes **aggregate statistics only**, not per-track audio data.

> ⚠️ **Provenance & usage.** This dataset contains transformative aggregate statistics derived from Spotify audio features and public artist metadata. The underlying audio features and metadata remain the property of **Spotify AB**. It is intended for **non-commercial research**. Not affiliated with or endorsed by Spotify.

---

## What is AVD?

AVD is a three-dimensional model of how people perceive music:

| Axis | Everyday meaning | Audio basis here |
|---|---|---|
| **A — Arousal** | Energy, intensity, force vs. calmness | Spotify `energy` |
| **V — Valence** | Positive, happy, bright vs. negative, sad, dark | Spotify `valence` |
| **D — Depth** | Sophistication, complexity, thoughtfulness, poetic quality | Derived from Spotify features (see below) |

The model comes from Greenberg et al. (2016), *The Song Is You*, who showed that listeners’ ratings of music collapse into these three components and that those components predict personality above and beyond demographics. Fricke et al. (2018) later demonstrated that the same AVD structure emerges from **computer-extracted** audio features, meaning the dimensions can be recovered automatically from the sound signal. That is what this dataset does at scale.

### How Depth is derived

Spotify does not ship a “Depth” feature. We construct it as an **acoustic-sophistication factor**:

1. Take the Spotify features that tap complexity and timbral sophistication: `acousticness`, `instrumentalness`, `danceability`, `speechiness`, `loudness`, and `tempo`.
2. **Residualise them against Arousal (`energy`) and Valence (`valence`)** so Depth is orthogonal to the other two axes.
3. Extract the first principal component, sign-check it against genre intuition (classical, ambient, jazz = high; comedy, drill, spoken-word = low), and rescale to 0–1.

The result is a single, interpretable axis that is largely uncorrelated with Arousal and Valence.

### Aggregation

Per-artist and per-genre values are **(popularity+1)-weighted means** over tracks, so an artist’s best-known sound dominates the profile. Each row also reports the standard deviation on each axis (`var_a`, `var_v`, `var_d`) and a composite `spread` score — the average standard deviation across AVD — which tells you how stylistically *varied* an artist or genre is.

Only tracks with a popularity of at least 5 enter the aggregates. This floor excludes the long tail of never-played catalogue entries (bootlegs, duplicates, mislabelled credits) that otherwise inflate track counts and spread without reflecting an artist’s actual sound.

---

## Files

### `artists_avd.parquet` — 2,204,210 artists

| column | type | meaning |
|---|---|---|
| `artist_key` | str | normalised (lowercased) artist name — join key |
| `artist_name` | str | display name |
| `artist_id` | str | Spotify artist id (base62) |
| `a`, `v`, `d` | float | Arousal, Valence, Depth on a 0–1 scale (popularity-weighted mean) |
| `var_a`, `var_v`, `var_d` | float | per-axis standard deviation across the artist’s tracks |
| `spread` | float | mean of the three std-devs — stylistic eclecticism/consistency |
| `n_tracks` | int | tracks with audio features used in the aggregate |
| `followers` | int | Spotify follower count |

### `genres_avd.parquet` — 734 genres

| column | type | meaning |
|---|---|---|
| `genre` | str | Spotify genre label |
| `n` | int | tracks used for the aggregate |
| `a`, `v`, `d` | float | Arousal, Valence, Depth (0–1) |
| `var_a`, `var_v`, `var_d` | float | per-axis standard deviation |
| `spread` | float | mean of the three std-devs |

Only genres with at least 10 tracks are included.

---

## A taste of the data

All values are on a 0–1 scale. The genre tables are restricted to genres with at least 1,000 tracks for stability; artist tables are restricted to artists with **>1 million followers** so the names are familiar.

### Genres — highest Depth

| genre | n tracks | a | v | d |
|---|---:|---:|---:|---:|
| ambient | 93,174 | 0.236 | 0.131 | **0.787** |
| classical piano | 209,837 | 0.109 | 0.275 | **0.785** |
| dark ambient | 61,675 | 0.230 | 0.124 | **0.777** |
| drone | 51,043 | 0.294 | 0.130 | **0.776** |
| space music | 34,534 | 0.275 | 0.148 | **0.760** |

### Genres — lowest Depth

| genre | n tracks | a | v | d |
|---|---:|---:|---:|---:|
| comedy | 71,028 | 0.597 | 0.509 | **0.156** |
| spoken word | 52,055 | 0.447 | 0.457 | **0.236** |
| brooklyn drill | 16,440 | 0.640 | 0.489 | **0.255** |
| new york drill | 23,099 | 0.629 | 0.479 | **0.259** |
| uk drill | 32,558 | 0.622 | 0.574 | **0.264** |

### Genres — highest Valence

| genre | n tracks | a | v | d |
|---|---:|---:|---:|---:|
| vallenato | 36,785 | 0.689 | **0.839** | 0.552 |
| duranguense | 27,315 | 0.658 | **0.825** | 0.569 |
| tropical music | 58,531 | 0.652 | **0.823** | 0.563 |
| merengue | 49,977 | 0.732 | **0.812** | 0.564 |
| cumbia | 219,370 | 0.707 | **0.809** | 0.561 |

### Genres — lowest Valence

| genre | n tracks | a | v | d |
|---|---:|---:|---:|---:|
| dark ambient | 61,675 | 0.230 | **0.124** | 0.777 |
| drone | 51,043 | 0.294 | **0.130** | 0.776 |
| ambient | 93,174 | 0.236 | **0.131** | 0.787 |
| space music | 34,534 | 0.275 | **0.148** | 0.760 |
| black metal | 51,649 | 0.815 | **0.165** | 0.698 |

### Genres — highest Arousal

| genre | n tracks | a | v | d |
|---|---:|---:|---:|---:|
| hardcore techno | 39,996 | **0.934** | 0.268 | 0.530 |
| deathcore | 48,999 | **0.924** | 0.221 | 0.571 |
| hardcore | 103,663 | **0.922** | 0.289 | 0.518 |
| frenchcore | 61,114 | **0.918** | 0.261 | 0.457 |
| gabber | 39,131 | **0.916** | 0.278 | 0.486 |

### Genres — lowest Arousal

| genre | n tracks | a | v | d |
|---|---:|---:|---:|---:|
| classical piano | 209,837 | **0.109** | 0.275 | 0.785 |
| chamber music | 172,209 | **0.131** | 0.289 | 0.718 |
| requiem | 104,235 | **0.139** | 0.224 | 0.679 |
| classical | 566,956 | **0.141** | 0.251 | 0.735 |
| choral | 140,617 | **0.174** | 0.210 | 0.652 |

### Artists (>1M followers) — highest Depth

| artist | followers | a | v | d |
|---|---:|---:|---:|---:|
| my bloody valentine | 1,039,825 | 0.718 | 0.320 | **0.832** |
| Frédéric Chopin | 3,470,884 | 0.089 | 0.183 | **0.799** |
| Max Richter | 1,077,258 | 0.134 | 0.107 | **0.795** |
| Ludovico Einaudi | 3,777,203 | 0.143 | 0.109 | **0.786** |
| Ludwig van Beethoven | 5,565,695 | 0.117 | 0.257 | **0.785** |

### Artists (>1M followers) — lowest Depth

| artist | followers | a | v | d |
|---|---:|---:|---:|---:|
| Die drei ??? | 1,142,355 | 0.486 | 0.542 | **0.082** |
| Shoreline Mafia | 1,301,140 | 0.560 | 0.371 | **0.179** |
| Ugly God | 2,045,624 | 0.485 | 0.419 | **0.192** |
| Blueface | 6,819,253 | 0.564 | 0.491 | **0.194** |
| King Von | 6,433,332 | 0.610 | 0.412 | **0.195** |

### Artists (>1M followers) — highest Valence

| artist | followers | a | v | d |
|---|---:|---:|---:|---:|
| Jorge Oñate | 1,051,332 | 0.683 | **0.945** | 0.579 |
| Diomedes Diaz | 5,667,222 | 0.672 | **0.935** | 0.593 |
| El Coyote Y Su Banda Tierra Santa | 1,688,623 | 0.579 | **0.928** | 0.616 |
| Binomio de Oro | 1,699,124 | 0.628 | **0.925** | 0.559 |
| Grupo Laberinto | 1,021,440 | 0.498 | **0.924** | 0.538 |

### Artists (>1M followers) — lowest Valence

| artist | followers | a | v | d |
|---|---:|---:|---:|---:|
| Claude Debussy | 1,607,806 | 0.048 | **0.096** | 0.777 |
| Sigur Rós | 1,395,389 | 0.353 | **0.107** | 0.744 |
| Max Richter | 1,077,258 | 0.134 | **0.107** | 0.795 |
| Ludovico Einaudi | 3,777,203 | 0.143 | **0.109** | 0.786 |
| Hans Zimmer | 4,470,974 | 0.289 | **0.111** | 0.760 |

### Artists (>1M followers) — highest Arousal

| artist | followers | a | v | d |
|---|---:|---:|---:|---:|
| Cannibal Corpse | 1,157,976 | **0.980** | 0.198 | 0.732 |
| Slayer | 4,260,332 | **0.966** | 0.183 | 0.578 |
| Lamb of God | 2,211,438 | **0.964** | 0.208 | 0.554 |
| Amon Amarth | 1,199,599 | **0.959** | 0.214 | 0.662 |
| Killswitch Engage | 2,063,594 | **0.953** | 0.297 | 0.566 |

### Artists (>1M followers) — lowest Arousal

| artist | followers | a | v | d |
|---|---:|---:|---:|---:|
| Claude Debussy | 1,607,806 | **0.048** | 0.096 | 0.777 |
| Franz Schubert | 1,166,930 | **0.081** | 0.180 | 0.724 |
| Frédéric Chopin | 3,470,884 | **0.089** | 0.183 | 0.799 |
| Johannes Brahms | 1,040,044 | **0.102** | 0.223 | 0.774 |
| Ludwig van Beethoven | 5,565,695 | **0.117** | 0.257 | 0.785 |

### Artists (>1M followers) — most eclectic vs. most consistent

`spread` is the average of `var_a`, `var_v`, and `var_d`. A high `spread` means the artist’s catalogue roams widely across AVD space; a low `spread` means the tracks cluster tightly around one signature. To keep “most consistent” meaningful, the consistency table requires at least 20 tracks.

#### Most eclectic

| artist | followers | n tracks | a | v | d | spread |
|---|---:|---:|---:|---:|---:|---:|
| Toby Fox | 1,730,867 | 466 | 0.509 | 0.584 | 0.757 | **0.262** |
| Jim Morrison | 1,896,305 | 52 | 0.422 | 0.432 | 0.455 | **0.244** |
| Salman Khan | 1,413,888 | 100 | 0.690 | 0.534 | 0.484 | **0.242** |
| Böhse Onkelz | 1,352,442 | 839 | 0.802 | 0.411 | 0.531 | **0.237** |
| Santhosh Narayanan | 3,087,709 | 965 | 0.569 | 0.476 | 0.476 | **0.236** |

#### Most consistent

| artist | followers | n tracks | a | v | d | spread |
|---|---:|---:|---:|---:|---:|---:|
| Eslam Kabonga | 1,956,555 | 658 | 0.886 | 0.794 | 0.401 | **0.065** |
| Hamo Bika | 2,353,706 | 404 | 0.883 | 0.815 | 0.427 | **0.070** |
| Os Barões Da Pisadinha | 6,826,354 | 507 | 0.875 | 0.913 | 0.669 | **0.073** |
| JISOO | 5,613,776 | 55 | 0.727 | 0.604 | 0.420 | **0.078** |
| Claude Debussy | 1,607,806 | 4,408 | 0.048 | 0.096 | 0.777 | **0.079** |

---

## Why this dataset is useful

Because AVD is grounded in music-psychology research, these aggregates are useful far beyond simple tagging:

- **Recommendation & discovery** — model similarity in a psychologically meaningful space instead of raw audio features.
- **Genre taxonomy** — compare genres on continuous emotion and complexity axes rather than folk labels.
- **Playlist generation** — build playlists that move through AVD space or stay within a narrow emotional pocket.
- **Artist similarity** — find artists with similar AVD signatures but different follower sizes, eras, or scenes.

---

## Quick start

```python
from datasets import load_dataset

artists = load_dataset("<user>/spotify-avd", "artists", split="train")
genres  = load_dataset("<user>/spotify-avd", "genres",  split="train")

df = artists.to_pandas()

# Most eclectic well-known artists
df[df.followers > 1_000_000].nlargest(10, "spread")[
    ["artist_name", "a", "v", "d", "spread", "followers"]
]

# Chill, complex, low-arousal genres
genres_df = genres.to_pandas()
genres_df.query("a < 0.3 and d > 0.7").nsmallest(10, "v")[
    ["genre", "a", "v", "d", "n"]
]
```

---

## Limitations

- **Arousal and Valence are Spotify’s estimates**, not ground-truth human judgments. They are convenient and well-validated proxies, but still black-box predictions.
- **Depth is a derived construct**, not a directly measured label. We sign-checked it against genre intuition, but there is no public large-scale Depth ground-truth dataset.
- The snapshot reflects the **2025-07 Spotify catalogue**; newer releases are absent.
- Aggregate means hide within-artist range — use `spread` and `var_*` when consistency matters.
- Spotify follower counts are a noisy proxy for prominence and vary over time.
- These are **musical-attribute profiles only**. They describe sound, not listeners, and carry no clinical or psychological-state meaning about any individual.

---

## Citation & references

If you use this dataset in research, please cite the underlying AVD work and this dataset:

```bibtex
@article{greenberg2016song,
  title={The Song Is You: Preferences for Musical Attribute Dimensions Reflect Personality},
  author={Greenberg, David M. and Kosinski, Michal and Stillwell, David J. and Monteiro, Brian L. and Levitin, Daniel J. and Rentfrow, Peter J.},
  journal={Social Psychological and Personality Science},
  volume={7},
  number={6},
  pages={597--605},
  year={2016}
}

@article{fricke2018computer,
  title={Computer-based music feature analysis mirrors human perception and can be used to measure individual music preference},
  author={Fricke, Kai R. and Greenberg, David M. and Rentfrow, Peter J. and Herzberg, Philipp Yorck},
  journal={Journal of Research in Personality},
  volume={75},
  pages={94--102},
  year={2018}
}
```

Dataset built with [Spotilyze](https://github.com/flaser381/spotilyze).
</content>
</invoke>