Sound Classifier

1. Perceptual Frequency Analysis:
   • Utilizes the Mel scale for filter design, emphasizing frequencies relevant to human auditory perception.
   • Gabor filters provide time-frequency localization, critical for capturing audio features effectively.

2. Efficient Feature Extraction:
   • Audio signals are segmented into overlapping windows.
   • Each window is processed with a Gabor filter bank to extract mean and standard deviation responses.
   • The resulting feature vector captures rich time-frequency information.

1. Simple yet Effective Classifier:
   • A KNN classifier is trained using the extracted features.
   • Designed for easy evaluation and modification.

Metrics

Metric	Value
Training Accuracy	77%
Testing Accuracy	67%

Performance aligns with the expected range of 55% - 68% for the test set.

Structure

• data (Dataset data.mat)

• images (Folder in which the images are saved)

• knn_audio.py

  • Loads audio data.
  • Extracts features using get_features.
  • Trains and evaluates a KNN classifier.

• include (All functions are included here)

  • [1] gabor_filters.py (Generates Gabor filters based on the Mel scale):
    • gabor_filter: Generates a Gabor filter (cosine and sine components).
    • gabor_filter_bank: Creates a bank of Gabor filters across a range of Mel-scale frequencies.
    • gaussian_filter: Creates Gaussian filters for comparison.
    • from_hz_to_mel, from_mel_to_hz: Converts between Hertz and Mel scale.
  • [2] get_features.py (Core features extractor):
    • Segments audio into overlapping windows.
    • Applies the Gabor filter bank to extract features (mean and standard deviation for each filter response).
    • Outputs a feature vector representing the signal.
  • [3] plots.py (Plotting tools):
    • Spectrum for Gaussian and Gabor filters.
    • Mel scale versus normal frequency mapping.
    • Time-frequency representations.