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

Area of Science:

• Computational linguistics
• Machine learning for healthcare
• Speech signal processing

Background:

• Depression detection from speech is vital for mental health assessment.
• Current methods suffer from poor feature extraction, interpretability issues, and language barriers.
• There is a need for more accurate and efficient depression detection models.

Purpose of the Study:

• To propose a novel speech-based depression detection model with enhanced accuracy and performance.
• To address the limitations of existing depression detection systems.
• To develop a robust model for identifying different stages of depression.

Main Methods:

• Adaptive threshold-based pre-processing (AdaT) and twinned Savitzky-Golay filter (TSaG) for noise reduction.
• Synchro-Squeezed Adaptive Wavelet Transform Algorithm (SSawT) for signal-to-image conversion.
• Singular Empirical Decomposition and Sparse Autoencoder (SiFE) for feature extraction.
• Weighted Soft Attention-based Fusion (WSAttF) for combining features.
• Chaotic Mud Ring Optimization (ChMR) for feature selection.
• Dilated Convolutional Neural Network (CNN) based Bidirectional-Long Short Term Memory-Bi-LSTM (DiCBiL) for depression stage detection.

Main Results:

• The proposed model achieved 93.22% F1-score, 93.11% precision, 93.12% recall, and 93.31% accuracy on the DAIC-WOZ test set.
• Validation on AVEC 2019 dataset yielded 93.91% accuracy.
• Validation on MELD dataset achieved 85.34% accuracy.
• The model demonstrates high effectiveness in detecting depression stages with reduced error rates.

Conclusions:

• The proposed speech-based depression detection model significantly improves accuracy and efficiency.
• The novel combination of signal processing, feature extraction, and deep learning techniques is effective.
• This model offers a promising tool for objective mental health assessment and depression diagnosis.