Research on Acoustic Scene Classification Based on Time-Frequency-Wavelet Fusion Network.

Area of Science:

• Artificial Intelligence
• Signal Processing
• Machine Learning

Background:

• Acoustic scene classification identifies environments from sound signals.
• Variations in audio due to different cities and devices challenge model accuracy.
• Existing methods struggle with diverse and complex acoustic environments.

Purpose of the Study:

• To develop an advanced acoustic scene classification model.
• To enhance model robustness against audio variations.
• To improve classification accuracy in real-world scenarios.

Main Methods:

• A time-frequency-wavelet fusion network was proposed.
• A time-frequency-wavelet module extracted features across time, frequency, and wavelet domains.
• Gated temporal-spatial attention and visual state space modules were integrated for enhanced contextual modeling.
• Kolmogorov-Arnold network layers replaced traditional multilayer perceptrons in the classifier.

Main Results:

• Achieved 56.16% average accuracy on the TAU Urban Acoustic Scenes 2022 mobile dataset, a 6.53% improvement over the baseline.
• Reached 97.60% accuracy on the UrbanSound8K dataset, outperforming existing methods.
• Demonstrated significant performance gains in complex acoustic scenarios.

Conclusions:

• The proposed time-frequency-wavelet fusion network effectively improves acoustic scene classification accuracy.
• The model exhibits strong generalization ability across different datasets and complex environments.
• The fusion of multidimensional audio information is crucial for robust acoustic scene recognition.