DATASET_CARD.md

# Dataset Card for HTMOneShotLoopClassification

## Dataset Description

- **Homepage**: https://github.com/itsuzef/HTMOneShotLoopClassification
- **Repository**: https://github.com/itsuzef/HTMOneShotLoopClassification
- **Paper**: [To be added - link to paper when published]
- **Point of Contact**: https://github.com/itsuzef
- **DOI**: [10.5281/zenodo.17872191](https://doi.org/10.5281/zenodo.17872191)

### Dataset Summary

HTMOneShotLoopClassification is a dataset of 5,561 electronic music samples (one-shots and loops) from house, tech house, and minimal techno genres. Each sample is labeled with one of 8 stem categories and includes 55 extracted audio features for machine learning applications.

The dataset is designed for:
- Audio classification tasks
- Stem category prediction
- Feature engineering research
- Electronic music analysis

### Supported Tasks and Leaderboards

- **Audio Classification**: Multi-class classification into 8 stem categories
- **Baseline Performance**: 85.41% accuracy using SVM with RBF kernel

### Languages

- Audio samples (no spoken language)
- Metadata in English

## Dataset Structure

### Data Instances

Each instance contains:
- **Filename**: Original audio file name
- **Stem Label**: One of 8 categories (kick, snare, hihat, bass, synth, vocal, percussion, fx)
- **Category**: Sample type (one-shots or loops)
- **55 Audio Features**: Extracted using librosa

### Data Fields

| Field | Type | Description |
|-------|------|-------------|
| filename | string | Original audio filename |
| stem | categorical | Target label (8 categories) |
| category | categorical | one-shots or loops |
| duration | float | Audio duration in seconds |
| sample_rate | integer | Sample rate (typically 44100 Hz) |
| rms | float | Root mean square energy |
| zcr | float | Zero crossing rate |
| spectral_centroid | float | Spectral centroid in Hz |
| ... | ... | (See metadata_schema.json for all 55 features) |

### Data Splits

Recommended splits:
- **Train**: 80% (4,450 samples)
- **Validation**: 10% (556 samples)
- **Test**: 10% (556 samples)

Note: Good class balance (FX: 561 samples, Hihat: 907 samples). All categories within 80-130% of average.

## Dataset Creation

### Curation Rationale

The dataset was created to address the lack of curated electronic music sample datasets with:
- High-quality professional samples
- Consistent feature extraction
- Genre-specific focus (house/tech house/minimal techno)
- Both one-shots and loops

### Source Data

- Professional electronic music sample packs
- Curated from house, tech house, and minimal techno collections
- Quality-controlled and deduplicated

**Category Notes**:
- **Hihat** (907 samples): Most represented category because it includes multiple cymbal types (closed hihats, open hihats, ride cymbals, crash cymbals) essential for rhythmic patterns
- **Bass & Synth** (100% key coverage): Key information is critical for these melodic categories to enable harmonic mixing and prevent dissonance
- **Vocal** (45.1% key coverage): Samples without keys are primarily spoken word, vocal chops, and non-melodic elements; melodic vocals have key information

### Annotations

- **Annotation process**: Manual categorization by stem type
- **Annotators**: Domain experts
- **Annotation guidelines**: Based on standard electronic music production categories

### Personal and Sensitive Information

No personal or sensitive information is included. All samples are from commercial sample packs or original productions.

## Considerations for Using the Data

### Social Impact of Dataset

This dataset enables research in:
- Automated music production tools
- Audio classification systems
- Electronic music analysis

### Discussion of Biases

- **Genre Bias**: Focused on house, tech house, and minimal techno (not all electronic music)
- **Class Balance**: Good balance achieved (FX: 561, all categories within 80-130% of average)
- **Key Coverage**: Not all categories have musical key information (only bass and synth have 100% coverage)

### Other Known Limitations

- Dataset size: 5,561 samples (moderate size)
- Limited to specific sub-genres of electronic music
- Some categories have low key coverage (important for harmonic mixing applications)

## Additional Information

### Dataset Curators

- Youssef Hemimy

### Licensing Information

**Dataset License**: CC-BY-4.0 (Creative Commons Attribution 4.0 International)

**License Scope**: This license applies to the CSV dataset containing extracted audio features and metadata, documentation, and scripts. You are free to use, share, and adapt the dataset for any purpose, including commercial use, provided you give appropriate credit.

**Original Audio Files**: The original audio files used to generate this dataset are NOT included and are NOT covered by this license. The audio files were obtained from commercial sample packs under their respective licenses, which typically prohibit redistribution. Users must obtain the original audio files separately if needed.

**What You Get**: This dataset provides extracted features (MFCCs, spectral features, etc.) and metadata that can be used for machine learning research without requiring access to the original audio files.

See `LICENSING.md` for complete licensing details.

### Citation Information

```bibtex
@dataset{hemimy2025htmoneshotloop,
  title={HTMOneShotLoopClassification: A Dataset of House, Tech House, and Minimal Techno Samples for Stem Category Classification},
  author={Hemimy, Youssef},
  year={2025},
  publisher={Zenodo},
  doi={10.5281/zenodo.17872191}
}
```

### Contributions

Contributions and feedback are welcome. Please see the repository for contribution guidelines.