Computational musicology project analyzing ZUN's Touhou Project soundtracks. Extracted 110+ audio features from 379 tracks across 19 games (TH01-TH19) to empirically measure compositional evolution, game atmospheres, and stage vs boss theme differences.

Live demo: touhou-web.vercel.app/explorer

Key insight: ZUN's music has gotten slower and moodier over 20 years. The data confirms what fans intuit.

Key insight: Boss themes emphasize weight over speed.

• PC-98: Happy/Excited quadrant (bright FM synthesis)
• Early Windows: High arousal, spans positive-negative
• Mid Windows: Shifted toward Tense (darker, energetic)
• Late Windows: Drifted to Sad/Calm (melancholic)

379 tracks × 19 games
        ↓
   [catalog.py] Scan OST folders
        ↓
   [enrich_catalog.py] Add metadata (era, stage position, character)
        ↓
   [features.py] Extract 110+ audio features via librosa
        ↓
   features.json (379 tracks × 110+ features)
        ↓
   Analysis notebooks (7 experiments)

• Spectral centroid, bandwidth, rolloff, flatness
• Zero-crossing rate

• Tempo (BPM), beat strength, onset rate

• Chroma (12 pitch classes), chroma entropy

• MFCCs (13 coefficients)
• Spectral entropy

Each time-varying feature gets: mean, std, min, max, p10, p50, p90, skew, range

conda create -n touhou-analysis python=3.11
conda activate touhou-analysis
pip install -r requirements.txt

# Run pipeline
python src/extraction/catalog.py        # Scan OST folders
python src/extraction/enrich_catalog.py # Add metadata
python src/extraction/features.py       # Extract features (~17 min)

# Run analyses
python notebooks/01_era_analysis.py
python notebooks/02_stage_position_analysis.py
# ... etc

touhou-composition-analysis/
├── data/
│   ├── raw/                    # Symlink to OST collection (FLAC)
│   ├── processed/
│   │   ├── features.json       # 379 tracks × 110+ features
│   │   └── umap_coords.json    # 2D projections for visualization
│   └── metadata/
│       └── catalog.json        # Enriched track metadata
├── src/
│   ├── extraction/             # Data pipeline
│   ├── analysis/               # DataFrame utilities
│   └── visualization/          # Plotting helpers
├── notebooks/                  # 7 analysis scripts
├── outputs/figures/            # Generated visualizations
└── DECISIONS.md                # Technical decision log

See DECISIONS.md for detailed rationale on:

• Why handcrafted features over deep learning
• Era boundary definitions
• Stage position classification
• PC-98 handling strategy
• Statistical methodology

• touhou-style-classifier: ML classifier for doujin circle arrangement style (89.5% accuracy on 5 circles)
• touhou-web: Interactive web demo with UMAP track explorer and personal game commentary

MIT