A CNN Sound Classification Mechanism Using Data Augmentation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Signal Processing

Background:

• Deep learning offers effective sound classification, surpassing traditional methods.
• Challenges in deep learning for sound include data imbalance, annotation issues, and resource constraints.
• Mel-Frequency Cepstral Coefficients (MFCCs) are used for feature extraction, converting sound into spectrograms suitable for CNN input.

Purpose of the Study:

• To propose a novel sound classification mechanism leveraging convolutional neural networks (CNNs).
• To enhance sound classification performance by implementing a data augmentation technique.
• To address data limitations and improve accuracy in complex sound datasets.

Main Methods:

• Utilized convolutional neural networks (CNNs) for sound classification.
• Employed Mel-Frequency Cepstral Coefficients (MFCCs) to transform audio signals into spectrograms.
• Implemented a data augmentation strategy by adjusting the number of triangular bandpass filters (K).

Main Results:

• Achieved a significant accuracy increase from 63% to 97% on the ESC-50 dataset using data augmentation (K=5).
• Demonstrated high accuracy (90%) on the UrbanSound8K dataset, further improved to 92% with augmentation.
• Showcased accelerated model training and 91% accuracy using 50% of the training data combined with augmentation.

Conclusions:

• The proposed data augmentation method substantially improves CNN-based sound classification accuracy, particularly for imbalanced or limited datasets.
• CNNs combined with MFCC feature extraction and effective data augmentation provide a robust solution for diverse sound classification tasks.
• Data augmentation accelerates model training and maintains high performance even with reduced training data, offering practical advantages.