Scientists Enhance Voice Pathology Detection through Adversarial Continual Learning

Vocal pathology refers to problems caused by abnormal conditions, such as dysphonia, paralysis, cysts, and even cancer, resulting in abnormal vibrations of the vocal cords (or vocal folds). In this context, Vocal Pathology Detection (VPD) has attracted widespread attention as a non-invasive method for automatically detecting voice problems. It consists of two processing modules: a feature extraction module for characterizing normal voice and a voice detection module for detecting abnormal voice.

Machine learning methods such as Support Vector Machines (SVM) and Convolutional Neural Networks (CNN) have been successfully applied to vocal pathology detection with the goal of the voice pathology detection module, achieving good VPD performance. In addition, pre-trained self-supervised models can learn general and rich representations of speech features, rather than explicit speech features, further improving their VPD capabilities.

However, fine-tuning these models for VPD can lead to overfitting problems due to domain shift from conversational speech to VPD tasks. As a result, pre-trained models become overly focused on training data and perform poorly on new data, failing to generalize.

To alleviate this problem, a group of researchers from the Gwangju Institute of Science and Technology (GIST) in Korea, led by Professor Hong Kook Kim, proposed a contrastive learning method involving Wave2Vec 2.0, which is a self-supervised pre-training model for speech signals, and adopted a new method called Adversarial Task Adaptive Pre-training (A-TAPT). They combined adversarial regularization during the continual learning process.

The researchers conducted various VPD experiments using the Saarbrucken Voice Database and found that the proposed A-TAPT improved the Unweighted Average Recall (UAR) by 12.36% and 15.38% compared to SVM and CNN ResNet50, respectively. It also outperformed traditional TAPT learning by 2.77% in UAR. This indicates that A-TAPT is more effective in mitigating overfitting problems.

Regarding the long-term impact of this work, Mr. Park, the first author of the paper, stated, "In the next five to ten years, our VPD research conducted in collaboration with MIT may fundamentally change healthcare, technology, and various industries. By achieving early and accurate diagnosis of voice-related diseases, it may bring more effective treatments and improve the quality of life for countless individuals."