Sound texture recognition through dynamical systems modeling of empirical mode decomposition

Area of Science:

• Signal Processing
• Machine Learning
• Acoustics

Background:

• Sound texture analysis is crucial for audio recognition and music information retrieval.
• Existing methods often struggle with the complexity and variability of natural sound textures.

Purpose of the Study:

• To develop and evaluate a novel system for modeling, recognizing, and classifying sound textures.
• To adapt techniques from video texture analysis for audio signal processing.

Main Methods:

• Empirical Mode Decomposition (EMD) for time/frequency analysis to represent signals as mode functions.
• Linear Dynamical System (LDS) modeling to capture the dynamics of these modes.
• Application of both linear and nonlinear techniques for learning system dynamics.

Main Results:

• The system achieved a 90% classification accuracy on a dataset of five distinct sound textures (fire, typewriter, rain, beverages, applause).
• The proposed LDS-based approach significantly outperformed a Mel-Frequency Cepstral Coefficient (MFCC) based LDS model.
• Performance also surpassed a standard Gaussian Mixture Model (GMM) classifier.

Conclusions:

• The developed system effectively models and classifies sound textures, demonstrating the efficacy of adapting video texture analysis methods to audio.
• The LDS approach, combined with EMD, provides a robust framework for distinguishing between complex sound textures.
• This research offers a promising new direction for audio analysis and sound recognition applications.