Noise-Resilient Audio Fingerprinting (CFAR + Inverted Index)

This project implements an audio fingerprinting pipeline that is robust to noise, based on:

• STFT-based anchor point selection with Advanced CFAR filtering (CA, OS, SO, TM)
• Pairwise hashing within target zones
• An inverted index for fast matching

What's new (refactor)

• Advanced CFAR Modes:
  • CA (Cell Averaging): Standard mode, best for Gaussian white noise (AWGN).
  • OS (Ordered Statistic): Uses 75th percentile, robust to impulsive noise (Nature recordings).
  • SO (Smallest Of): Uses min(left_window, right_window), best for dense music (avoids masking weak signals near strong beats).
  • TM (Trimmed Mean): Removes top 20%/bottom 10% outliers, offering a robust middle-ground.
• Refactored Structure:

Algorithm Performance Comparison

Below are the experimental results comparing different CFAR algorithms across two major noise scenarios.

1. Scenario: [NATURE] Natural Noise

Query: GeorgeDataset (10s snippets) | Reference: GTZAN (999 songs)

Algorithm	Accuracy
TM-CFAR	79.04%
SO-CFAR	76.20%
OS-CFAR	75.83%
CA-CFAR	72.05%

2. Scenario: [AWGN] Gaussian White Noise

Condition: SNR = 0dB, Duration = 5s (Extremely challenging)

Algorithm	Accuracy
TM-CFAR	51.00%
SO-CFAR	48.50%
CA-CFAR	45.50%
OS-CFAR	44.00%
OFF (No Filter)	16.00%

• Refactored Structure:
  • audiofp/: Core package (fingerprint, index).
  • main.py: Unified entry point with easy Scenario/CFAR configuration.
• Multiprocessing: Enabled by default for faster index building.

Project structure

.
├── audiofp
│   ├── __init__.py
│   ├── fingerprint.py
│   └── index.py
├── main.py
├── Data/
│   └── GTZAN/                         # your dataset folder (example)
└── Inverted_Index/
    └── GTZAN_STFT_inverted_index_table.pkl   # where you save pickles

Installation

Python 3.9+ recommended.

pip install numpy librosa matplotlib seaborn scikit-learn tqdm

Quick start

1. Configure Experiment in main.py

Open main.py and edit the configuration block at the bottom:

if __name__ == "__main__":
    # ==========================================
    #             EXPERIMENT CONFIG
    # ==========================================
    
    # Select Scenario: "AWGN" or "NATURE"
    SCENARIO = "NATURE" 

    # Select CFAR Algorithm: "CA", "OS", "SO", "TM"
    CFAR_MODE = "TM"  

2. Run the script

python main.py

It will automatically select the appropriate dataset paths (if configured) and run the matching experiment, printing the classification accuracy.

3. Use as a library (recommended)

from audiofp import MusicFingerprint, inverted_index_table, music_to_folder_matching

# Build inverted index
ii = inverted_index_table("Data/GTZAN/", multiprocess=True, CFAR_mode="CA")

# Query with specific CFAR mode
best, counts, deltas = music_to_folder_matching(
    music_path="Data/query_snippet.wav",
    music_name="query.wav",
    folder_path="",
    inverted_index=ii,
    CFAR_mode="TM",  # Use Trimmed Mean for query
)

print("Best match:", best)

Background & Inspiration

This project originates from a Master's course project in Music Informatics.

The core innovation stems from the author's background in Radar Signal Processing. In radar systems, detecting a target against a complex background (clutter) is a classic problem, often solved using CFAR (Constant False Alarm Rate) algorithms. These algorithms dynamically adjust the detection threshold based on local noise statistics to maintain a stable false alarm rate.

Inspired by this, this project treats:

• Audio Spectrogram Peaks as "radar targets".
• Background Music/Noise as "clutter".

By cross-applying radar technology to audio information retrieval, we implement and evaluate multiple CFAR variants to robustly extract audio fingerprints under challenging conditions:

• CA-CFAR: The classic baseline, effective for uniform noise.
• OS/SO/TM-CFAR: Advanced variants designed to handle non-homogeneous acoustic environments, impulsive noise, and dense polyphonic textures.

Authors

Ziyue Yang, Yuqi Zhang

License

MIT