A novel dilated Bi-LSTM framework for depression detection from speech signals through feature fusion

A crucial method for determining a person's mental health and assessing their degree of depression is depression detection. To identify depression through speech or conversation, a number of sophisticated methods and questionnaires have been created. The constraints of the current system are as follows: reduced effectiveness as a result of poor feature selection and extraction, problems with interpretability, and the difficulty of identifying depression in different languages. As a result, the proposed model is presented to offer improved accuracy and efficient performance. While adaptive threshold-based pre-processing (AdaT) is used to eliminate quiet and unnecessary information, the twinned Savitzky-Golay filter (TSaG) is used to minimize noise in the dataset. To turn the signal into an image, a Synchro-Squeezed Adaptive Wavelet Transform Algorithm (SSawT) is employed. The Singular Empirical Decomposition and Sparse Autoencoder (SiFE) model is used to extract linear and deep features. Input's deep, linear, and statistical properties are combined using the Weighted Soft Attention-based Fusion (WSAttF) model. From the fused features, the Chaotic Mud Ring Optimization algorithm (ChMR) chooses the best features. A Dilated Convolutional Neural Network (CNN) based Bidirectional-Long Short Term Memory-Bi-LSTM (DiCBiL) is used to detect different stages of depression, which lowers error rates and increases detection accuracy. The proposed method achieves 93.22% of F1-score, 93.11% precision, 93.12% recall, and 93.31% accuracy on the DAIC-WOZ original test set. During the testing time, two more datasets, namely AVEC 2019 and MELD, are used to validate the proposed performance, attaining an accuracy of 93.91% and 85.34% respectively.

The online version contains supplementary material available at 10.1007/s11571-026-10411-9.