Associated with Research Internship at the University of Victoria under the guidance of Prof.Geroge Tzanetakis
This repository presents a comparative benchmarking study of popular Music Information Retrieval (MIR) frameworks for beat tracking and tempo estimation tasks.
The project evaluates algorithmic performance across two widely used benchmark datasets:
- GTZAN Genre Collection
- GiantSteps Tempo Dataset
Using standardized evaluation metrics and statistical analysis, the study investigates the strengths, limitations, and dataset-dependent behavior of MIR algorithms under diverse musical conditions.
- Evaluate beat tracking performance across multiple datasets.
- Compare tempo estimation accuracy using established MIR frameworks.
- Analyze model behavior across different musical genres and styles.
- Perform statistical evaluation using standard MIR metrics.
- Assess the generalizability of algorithm performance across datasets.
A benchmark dataset containing audio recordings from ten music genres.
Genres include:
- Blues
- Classical
- Country
- Disco
- Hip-Hop
- Jazz
- Metal
- Pop
- Reggae
- Rock
Used for:
- Beat tracking evaluation
- Tempo estimation benchmarking
- Genre-wise analysis
A large-scale electronic dance music dataset with high-quality tempo annotations.
Used for:
- Tempo estimation benchmarking
- Cross-dataset generalization analysis
- Performance validation on modern electronic music
Existing research in Music Information Retrieval was reviewed to identify widely adopted beat tracking and tempo estimation approaches.
Algorithms were evaluated under identical experimental conditions and compared using standard MIR metrics.
Results were aggregated and analyzed using descriptive statistics and visualization techniques.
Performance was analyzed:
- Across datasets
- Across genres
- Across evaluation metrics
- Across algorithm implementations
The following metrics are used throughout the study:
| Metric | Purpose |
|---|---|
| F-measure | Beat tracking accuracy |
| Cemgil Score | Temporal precision |
| P-score | Beat sequence consistency |
| Information Gain | Beat prediction quality |
| Tempo Accuracy | Tempo estimation performance |
Madmom consistently outperformed Librosa across most musical genres, with particularly strong gains on Jazz, Pop, and Hip-Hop recordings.
.
├── notebooks/
│ ├── GTZAN_Benchmark.ipynb
│ └── GiantSteps_Benchmark.ipynb
│
├── figures/
│ ├── gtzan/
│ └── giantsteps/
│
├── README.md
├── requirements.txt
├── LICENSE
└── .gitignore
- Python
- Google Colab
- NumPy
- Pandas
- Matplotlib
- Seaborn
- librosa
- madmom
- mir_eval
- Comparative benchmarking across multiple MIR datasets.
- Reproducible evaluation pipeline.
- Statistical analysis of beat tracking and tempo estimation performance.
- Cross-dataset assessment of algorithm robustness.
- Research-oriented workflow suitable for future MIR studies.
- Downbeat tracking evaluation.
- Additional MIR datasets.
- Deep learning-based beat tracking models.
- Real-time tempo estimation systems.
- Cross-cultural music benchmarking.
- GTZAN Genre Collection
- GiantSteps Dataset
- librosa Documentation
- madmom Documentation
- mir_eval Evaluation Framework
Harshita Pulavarti
Research project exploring benchmarking methodologies for Music Information Retrieval systems, with a focus on beat tracking and tempo estimation.