Pitch Detection Benchmark

A comprehensive benchmark suite evaluating pitch detection algorithms across 8 datasets covering speech, music, synthetic, and real-world audio conditions.

Which Algorithm Should I Use?

TL;DR Recommendations:

• Best Overall: SwiftF0 (90.2% accuracy, 90× faster than CREPE)
• Need Maximum Speed: Praat (2.8ms per second of audio, 84.7% accuracy)
• Best Pitch Accuracy: CREPE (85.3% accuracy, best RPA/RCA but slow and not good on all metrics)
• Best Human singing: RMVPE (87.2% accuracy, best on Vocadito and MIR-1K)

Overall Results

The table below shows the harmonic-mean accuracy score for each algorithm across the eight benchmark datasets. The average score determines the overall ranking.

Algorithm	Bach10Synth	MDBStemSynth	MIR1K	NSynth	PTDB	PTDBNoisy	SpeechSynth	Vocadito	Average
SwiftF0	97.5%	92.0%	95.0%	89.3%	90.4%	74.0%	90.7%	92.6%	90.2%
RMVPE	98.1%	90.6%	96.0%	68.2%	88.9%	68.5%	90.6%	96.4%	87.2%
CREPE	98.5%	90.5%	95.7%	80.2%	79.7%	53.8%	88.3%	95.6%	85.3%
PENN	97.3%	94.0%	89.0%	63.3%	91.0%	76.4%	84.8%	82.4%	84.8%
Praat	96.0%	90.7%	92.6%	70.7%	86.2%	65.3%	88.2%	88.2%	84.7%
SPICE	95.0%	89.4%	92.7%	68.8%	77.8%	55.9%	87.9%	92.3%	82.5%
TorchCREPE	96.7%	85.1%	71.4%	83.8%	78.3%	61.2%	79.7%	89.0%	80.6%
pYIN	97.5%	90.3%	91.2%	74.3%	72.1%	43.2%	81.4%	79.5%	78.7%
RAPT	91.9%	79.6%	82.4%	54.6%	68.4%	48.9%	74.3%	87.5%	73.5%
SWIPE	77.8%	65.6%	77.1%	51.4%	66.6%	45.0%	77.1%	66.6%	65.9%
YAAPT	58.5%	39.6%	82.0%	6.4%	69.8%	51.7%	83.5%	88.6%	60.0%
BasicPitch	23.7%	12.4%	36.5%	77.7%	23.1%	12.6%	61.2%	17.8%	33.1%

For a detailed breakdown of results, see Benchmark Report.

Running Your Own Benchmarks

Installation

This project uses uv (a fast Python package manager) for dependency management, but conda or pip will also work.

uv venv --python 3.10
source .venv/bin/activate
uv pip install -r requirements.txt --extra-index-url https://download.pytorch.org/whl/cu126 --index-strategy unsafe-best-match

Dataset Setup

Download the required datasets:

• PTDB-TUG - Speech with laryngograph ground truth
• NSynth - Synthetic musical instruments
• MDB-stem-synth - Synthetic music stems
• MIR-1K - Vocal excerpts
• Vocadito - Solo vocal recordings
• Bach10-mf0-synth - Synthetic Bach compositions
• CHiME-Home - Background noise for testing

Organize datasets in a directory structure like:

datasets/
├── PTDB/
├── NSynth/
├── MDBStemSynth/
├── MIR1K/
├── Vocadito/
├── Bach10Synth/
└── chime_home/

Usage

1. Visualize Algorithms on Your Audio

python visualize_algorithms.py your_audio.wav --selected_algorithms SwiftF0 CREPE Praat

2. Speed Benchmark

python speed_benchmark.py --signal-length 1.0 --n-runs 20

3. Pitch Benchmark

for dataset in PTDB NSynth MIR1K Vocadito MDBStemSynth Bach10Synth; do
  python pitch_benchmark.py \
    --dataset $dataset \
    --data-dir datasets/$dataset \
    --chime-dir datasets/chime_home
done
python pitch_benchmark.py --dataset PTDBNoisy --data-dir datasets/PTDB --chime-dir datasets/chime_home
python pitch_benchmark.py --dataset SpeechSynth --data-dir datasets/speechsynth.pt --chime-dir audio_datasets/chime_home

4. Generate Report

python generate_report.py --results-dir results/ --output benchmark_report.md

Algorithm Implementations

The benchmark includes implementations of these algorithms:

Neural Networks:

• SwiftF0 - Fast CNN-based pitch detection
• CREPE - CNN-based pitch estimation
• TorchCREPE - PyTorch CREPE implementation
• PENN - Pitch-Estimating Neural Networks
• BasicPitch - Spotify's multi-instrument pitch detector
• SPICE - Self-supervised pitch estimation
• RMVPE - A Robust Model for Vocal Pitch Estimation in Polyphonic Music

Classical Methods:

• Praat - Autocorrelation-based
• pYIN - Probabilistic YIN
• YAAPT - Yet Another Algorithm for Pitch Tracking
• RAPT - Robust Algorithm for Pitch Tracking
• SWIPE - Sawtooth Waveform Inspired Pitch Estimator

🤝 Contributing

Contributions are welcome! To add a new algorithm, you can either submit a Pull Request with your own implementation or create an Issue to request it, and I will run the benchmark for you.

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

📚 Citation

If you use this benchmark in your research, please cite:

@misc{nieradzik2025swiftf0,
      title={SwiftF0: Fast and Accurate Monophonic Pitch Detection},
      author={Lars Nieradzik},
      year={2025},
      eprint={2508.18440},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2508.18440},
}