Depression detection from speech data using deep learning-based optimized temporal-frequency-channel attention with...

Area of Science:

• Artificial Intelligence
• Speech Signal Processing
• Clinical Psychology

Background:

• Detecting depression from voice is challenging due to subtle acoustic cues and poor cross-lingual generalization of current models.
• Existing methods often require transcriptions or visual data, limiting their applicability.
• Individual variations in speech patterns complicate accurate depression detection.

Purpose of the Study:

• To develop a lightweight, interpretable deep learning framework for depression detection directly from raw speech audio.
• To overcome limitations of cross-lingual generalization and reliance on transcriptions.
• To enhance model robustness and adaptability to diverse acoustic conditions.

Main Methods:

• A streamlined ResNet-18 model enhanced with a Temporal-Frequency-Channel Attention (TFCA) unit processes speech spectrograms.
• Raw audio is segmented into clips and converted to time-frequency representations.
• A novel Parameter Optimization with Conscious Allocation using Iterative Intelligence (POCAII) strategy optimizes hyperparameters for faster convergence and robustness.

Main Results:

• Achieved 89.38% segment-level accuracy and 93.94% subject-level accuracy on the DAIC-WOZ dataset.
• Reached 89.96% segment-level accuracy and 93.23% subject-level accuracy on the Androids Corpus.
• Demonstrated high segment-level Area Under the Receiver Operating Characteristic Curve (AUC) of 95.3% and 95.7% respectively, with interpretable attention visualizations.

Conclusions:

• The proposed deep learning framework effectively detects depression from speech audio with high accuracy and interpretability.
• The model shows strong cross-lingual and cross-dataset generalization capabilities.
• This approach offers a promising, transcription-free solution for scalable depression screening.