Audio enhancement and intelligent classification of household sound events using a sparsely deployed array

Area of Science:

• Acoustics and Signal Processing
• Machine Learning and Artificial Intelligence
• Audio Engineering

Background:

• Household sound event classification is crucial for smart environments.
• Existing systems often struggle with noisy environments and complex soundscapes.
• Microphone arrays offer potential for enhanced audio capture and processing.

Purpose of the Study:

• To develop and evaluate a robust household sound event classification system.
• To compare different audio localization methods for sparse microphone arrays.
• To assess the impact of signal enhancement on classification accuracy.

Main Methods:

• A cascaded system with an audio localization/enhancement front-end and a CNN-based classification back-end was designed.
• Localization methods compared include two-stage (subspace, constrained least squares) and direct (beamforming, MUSIC) approaches.
• Signal enhancement utilized a minimum variance distortionless response beamformer and a minimum-mean-square-error postfilter.
• Convolutional Neural Networks (CNNs) and Convolutional Long Short-Term Memory networks processed Mel-spectrograms for classification.

Main Results:

• Simulations and live room experiments evaluated the performance of direct and two-stage localization methods.
• The array-based front-end demonstrated significant signal quality enhancement.
• Enhanced audio processing led to improved sound event classification accuracy compared to a single microphone setup.

Conclusions:

• The proposed system effectively classifies household sound events.
• Sparse microphone arrays combined with signal enhancement improve classification performance.
• The developed system offers a promising solution for intelligent audio monitoring.