Self-Supervised Learning for Speech-Based Detection of Depressive States

Abstract

This study aims to enhance the accuracy of depression detection by leveraging representation learning from audio data. The data of depression speech sets are sparse and costly to annotate. Therefore, a self-supervised pre-training approach is employed to improve the performance, generalization capability, and training efficiency of downstream tasks. When processing unlabeled data, the pre-trained audio representations based on self-supervised learning may be interfered with by noisy data if there is a significant amount of noise or errors present. Consequently, it is necessary to effectively analyze long-distance sequence data to enhance anti-interference capabilities. However, traditional LSTM models have limitations in context extraction and robustness to input outliers. Thus, an improved method named CNN-BiLSTM is proposed in this paper. The network initializes the LSTM's embedding layer with pre-trained word vectors and extracts spatial and temporal features separately to ensure a full and complete expression of useful input information. Different weights are assigned based on the importance of the features to obtain fused features. Additionally, a random forest is used for classification to mitigate the risk of overfitting and to demonstrate good performance when processing high-dimensional data. Experimental results show that the proposed model exhibits good classification performance on the depression dataset, outperforming traditional methods and state-of-the-art investigations.