Transformers for Urban Sound Classification-A Comprehensive Performance Evaluation

Area of Science:

• Machine Learning
• Acoustic Event Detection
• Urban Sound Classification

Background:

• Urban environments generate numerous relevant sound events requiring accurate and timely classification.
• Existing models need improvement for effectiveness and promptness in identifying and duration-tracking sound events.
• Robust sound classification models are crucial for monitoring and analyzing urban acoustic scenes.

Purpose of the Study:

• To identify the best-performing model for classifying a wide range of urban sound events.
• To analyze and model Transformer architectures for urban sound event classification.
• To investigate the impact of pre-training, data augmentation, and complementary methods on model performance.

Main Methods:

• Extensive analysis and modeling of Transformer models on public urban sound datasets.
• Comparison of Transformer models against baseline and convolutional neural network (CNN) models.
• Evaluation of pre-training strategies (image and sound domains) and data augmentation techniques.

Main Results:

• A Transformer model utilizing a novel Adam optimizer with weight decay and AudioSet transfer learning achieved superior performance.
• High accuracy scores were recorded: 89.8% on UrbanSound8K, 95.8% on ESC-50, and 99% on ESC-10.
• Pre-training from the audio domain and data augmentation significantly improved model robustness.

Conclusions:

• Transformer models, particularly when combined with optimized training and transfer learning, represent a highly promising approach for urban sound classification.
• The proposed method offers effective and prompt identification of urban sound events, addressing key requirements for real-world applications.
• Best practices include leveraging pre-trained models and data augmentation for enhanced sound classification system performance.