EnViTSA: Ensemble of Vision Transformer with SpecAugment for Acoustic Event Classification

Area of Science:

• Artificial Intelligence
• Machine Learning
• Signal Processing

Background:

• Deep learning models for Acoustic Event Classification (AEC) face overfitting challenges due to high complexity.
• Existing methods struggle with data scarcity and generalization in acoustic event detection.

Purpose of the Study:

• To introduce EnViTSA, an innovative approach for Acoustic Event Classification.
• To enhance AEC performance by mitigating overfitting and addressing data scarcity.

Main Methods:

• Ensemble of pre-trained Vision Transformers combined with SpecAugment data augmentation.
• Transformation of raw acoustic signals into Log Mel-spectrograms.
• Application of time and frequency masking via SpecAugment to generate synthetic training data.

Main Results:

• Achieved 93.50% accuracy on ESC-10, 85.85% on ESC-50, and 83.20% on UrbanSound8K.
• Demonstrated significant reduction in overfitting through the ensemble approach.
• Validated the effectiveness of SpecAugment for AEC data augmentation.

Conclusions:

• EnViTSA offers a substantial advancement in Acoustic Event Classification.
• Vision Transformers and SpecAugment show strong potential for acoustic domain applications.
• The proposed method effectively tackles key challenges in deep learning for AEC.