---
dataset_info:
  features:
  - name: mel
    dtype:
      array2_d:
        shape:
        - 128
        - 300
        dtype: float32
  - name: label
    dtype:
      class_label:
        names:
          '0': Electronic
          '1': Experimental
          '2': Folk
          '3': Hip-Hop
          '4': Instrumental
          '5': International
          '6': Pop
          '7': Rock
  - name: track_id
    dtype: int64
  - name: artist_id
    dtype: int64
  - name: genre
    dtype: string
  splits:
  - name: train
    num_bytes: 15282641587
    num_examples: 99140
  - name: validation
    num_bytes: 3824051743
    num_examples: 24807
  - name: test
    num_bytes: 4810471185
    num_examples: 31206
  download_size: 23678275603
  dataset_size: 23917164515
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
license: apache-2.0
task_categories:
- audio-classification
language:
- en
tags:
- music
- audio
- fma
- genre-classification
- mel-spectrogram
pretty_name: FMA-Small Log-Mel-Spectrograms
size_categories:
- 100K<n<1M
---
# FMA - Small: Pre - computed Log - Mel - Spectrograms for Music Genre Classification

Pre - processed [FMA - Small](https://github.com/mdeff/fma) dataset containing **155,153** log - mel - spectrogram segments ready for training audio genre classifiers. No audio decoding needed - load and train directly.

## Dataset Details

| Property | Value |
|---|---|
| Source | FMA-Small (8,000 tracks × 30s) |
| Representation | Log - Mel - Spectrogram |
| Sample Shape | `(128, 300)` - 128 mel bins × 300 time frames |
| Sample Rate | 32,000 Hz |
| Segment Duration | 3 seconds (1.5s overlap → ~19 segments/track) |
| Classes | 8 genres |
| Split Strategy | `StratifiedGroupKFold` on `artist_id` (zero artist leakage) |

## Features

| Column | Type | Description |
|---|---|---|
| `mel` | `Array2D(float32)` `(128, 300)` | Log - mel - spectrogram segment |
| `label` | `ClassLabel` | Genre label `[0–7]` |
| `track_id` | `int64` | FMA track identifier |
| `artist_id` | `int64` | FMA artist identifier |
| `genre` | `string` | Human - readable genre name |

## Labels

`0` Electronic · `1` Experimental · `2` Folk · `3` Hip - Hop · `4` Instrumental · `5` International · `6` Pop · `7` Rock

## Splits

| Split | Samples | Ratio |
|---|---|---|
| `train` | 99,140 | ~64% |
| `validation` | 24,807 | ~16% |
| `test` | 31,206 | ~20% |

**No artist appears in more than one split** - enforced via `StratifiedGroupKFold` on `artist_id` to prevent data leakage.

## Quick Start

```python
from datasets import load_dataset

dd = load_dataset("minhqng/fma-small")
dd.set_format("torch", columns=["mel", "label"])

sample = dd["train"][0]
mel = sample["mel"]      # (128, 300) float32
label = sample["label"]  # int, 0–7
```

## Audio Processing Pipeline

Each 30 - second MP3 track was processed as follows:

```
MP3 → decode (PyAV) → mono → resample 32kHz → segment 3s (50% overlap)
    → MelSpectrogram (n_fft=1024, hop=320, 128 bins, Slaney norm)
    → log(mel + 1e-9) → truncate to 300 frames → (128, 300)
```

Silent/corrupt tracks and segments were removed before dataset creation.

## Citation

If you use this dataset, please cite the original FMA paper:

```bibtex
@inproceedings{defferrard2017fma,
  title = {{FMA}: A Dataset for Music Analysis},
  author = {Defferrard, Micha\"el and Benzi, Kirell and Vandergheynst, Pierre and Bresson, Xavier},
  booktitle = {18th International Society for Music Information Retrieval Conference (ISMIR)},
  year = {2017},
}
```