Acoustic scene classification based on three-dimensional multi-channel feature-correlated deep learning networks.

Area of Science:

• Signal Processing
• Machine Learning
• Artificial Intelligence

Background:

• Acoustic scene classification (ASC) is crucial for environmental perception.
• Distinguishing subtle environmental sound differences presents a significant challenge.

Purpose of the Study:

• To propose a novel approach for accurate acoustic scene classification.
• To improve the performance of environmental sound recognition systems.

Main Methods:

• Developed a multi-branch 3D CNN model for spatial-temporal feature extraction.
• Utilized expert acoustic knowledge and discrete wavelet transformations (DWT) for multi-frequency representations.
• Introduced a 3D CNN with residual connections and squeeze-and-excitation attentions (3D-SE-ResNet).
• Incorporated an auxiliary supervised branch using signal chromatograms to mitigate overfitting.

Main Results:

• The proposed multi-input, multi-feature 3D-CNN architecture demonstrated noticeable performance gains.
• Achieved superior results compared to state-of-the-art methods on the DCASE 2019 dataset.
• Effectively captured long-term and short-term correlations in environmental sounds.

Conclusions:

• The novel 3D-SE-ResNet architecture with multi-feature aggregation enhances ASC accuracy.
• The approach effectively addresses the challenges of subtle sound differences in environmental perception.
• The method shows significant potential for real-world acoustic scene analysis.