An Incremental Class-Learning Approach with Acoustic Novelty Detection for Acoustic Event Recognition.

Area of Science:

• Signal Processing
• Machine Learning
• Artificial Intelligence

Background:

• Acoustic scene analysis (ASA) is crucial for understanding complex sound environments.
• Traditional ASA struggles with non-stationary sounds and novel acoustic events, leading to performance degradation.
• Detecting and learning new sounds without prior knowledge is a significant challenge in audio analysis.

Purpose of the Study:

• To present a self-learning ASA framework for acoustic event recognition (AER).
• To enable the detection and incremental learning of novel acoustic events.
• To address the issue of catastrophic forgetting in machine learning models for audio analysis.

Main Methods:

• A six-element framework including signal pre-processing, feature extraction (low-level and deep learning models like VGG, ResNet, TDNN, TDNN-LSTM), acoustic novelty detection (AND), signal augmentation, and incremental class-learning (ICL).
• Self-supervised learning on extracted audio features without human intervention.
• Pre-training deep audio representations using the large-scale Google AudioSet dataset.

Main Results:

• The proposed framework successfully detects and incrementally learns novel acoustic events.
• Performance was validated using Mel-spectrograms and deep features on benchmark datasets (ESC-10, ESC-50, US8K) and a custom domestic environment dataset.
• The integration of AND and ICL demonstrated effective handling of new audio events.

Conclusions:

• The self-learning ASA framework offers a robust solution for acoustic event recognition in dynamic environments.
• The method effectively tackles catastrophic forgetting, allowing continuous learning of new sounds.
• This approach enhances the adaptability and performance of audio analysis systems when encountering unforeseen acoustic events.