Long-term scalogram integrated with an iterative data augmentation scheme for acoustic scene classification.

Area of Science:

• Signal Processing
• Machine Learning
• Acoustics

Background:

• Acoustic Scene Classification (ASC) relies on acoustic features for scene information extraction.
• Limited datasets can cause biased models, leading to poor performance on unseen data and ambiguous scene classes.

Purpose of the Study:

• To propose a novel long-term wavelet feature for enhanced ASC.
• To develop a data augmentation scheme for improved model generalization.

Main Methods:

• Extraction of a long-term wavelet feature (scalogram) for capturing discriminative scene information.
• Implementation of a data augmentation strategy using Auxiliary Classifier Generative Adversarial Networks (ACGANs) and a deep learning-based sample filter.
• Evaluation on Detection and Classification of Acoustic Scenes and Events (DCASE) datasets.

Main Results:

• The proposed wavelet feature (scalogram) offers lower storage requirements, faster classification, and higher accuracy than traditional Mel filter bank coefficients (FBank).
• The ACGAN-based data augmentation achieved a significant 6.10% absolute accuracy improvement on recordings from unseen cities.
• Experimental results on DCASE17 and DCASE19 datasets demonstrated a performance boost over FBank classifiers.

Conclusions:

• The proposed long-term wavelet feature is effective for ASC, providing efficiency and accuracy gains.
• The ACGAN-based data augmentation significantly improves the generalization capability of ASC systems, especially for diverse acoustic environments.