Efficient Noisy Sound-Event Mixture Classification Using Adaptive-Sparse Complex-Valued Matrix Factorization and OvsO...

Area of Science:

• Signal Processing
• Machine Learning
• Audio Analysis

Background:

• Sound event classification from noisy mixtures is challenging.
• Traditional methods often struggle with single-channel, noisy audio data.
• Accurate separation and classification are crucial for many audio applications.

Purpose of the Study:

• To propose a novel method for sound event classification from a single noisy audio mixture.
• To enhance sound event separation using an adaptive L1 sparsity Complex Nonnegative Matrix Factorization (CMF).
• To improve classification accuracy using a Support Vector Machine (SVM) with a mean supervector.

Main Methods:

• Extended Complex Nonnegative Matrix Factorization (CMF) with adaptive L1 sparsity for signal decomposition.
• Feature extraction including Mel frequency cepstral coefficients, short-time energy, and zero-crossing rate.
• Support Vector Machine (SVM) based one versus one (OvsO) strategy with mean supervector for classification.

Main Results:

• The proposed adaptive L1 sparsity CMF effectively decomposes noisy single-channel mixtures.
• The SVM-based OvsO classifier accurately categorizes demixed sound events.
• Experimental results demonstrate superior performance compared to state-of-the-art separation methods.

Conclusions:

• The integrated approach of adaptive L1 sparsity CMF and SVM-based classification offers a robust solution for sound event analysis in noisy conditions.
• The method shows significant improvements in both sound event separation and classification accuracy.
• This technique provides a promising advancement for real-world audio processing tasks.