Deep spectrotemporal network based depression severity estimation from speech.

Area of Science:

• Computational linguistics
• Psychiatry
• Machine learning

Background:

• Depression diagnosis is challenging due to reliance on subjective clinical evaluations.
• Objective methods for assessing depression severity are needed to improve early intervention.
• Speech analysis offers a promising avenue for automated depression detection.

Purpose of the Study:

• To develop a novel deep spectrotemporal network for estimating depression severity from vocal cues.
• To enhance the accuracy and objectivity of depression severity assessment.
• To improve upon existing machine learning approaches for speech-based depression analysis.

Main Methods:

• Utilized EfficientNet-B3 for extracting spectral features from Mel spectrograms.
• Introduced a novel Volume Local Neighborhood Encoded Pattern (VLNEP) descriptor for spatiotemporal feature capture.
• Employed a dual-stream transformer model for effective fusion of spectral and spatiotemporal features.

Main Results:

• The proposed deep spectrotemporal network demonstrated superior performance in estimating depression severity.
• Achieved state-of-the-art results on the AVEC2013 and AVEC2014 benchmark datasets.
• The novel feature extraction and fusion techniques significantly improved accuracy.

Conclusions:

• The developed framework offers a robust and accurate method for automated depression severity estimation using speech.
• This approach has the potential to aid clinicians in objective and timely diagnosis of depression.
• Further research can explore broader applications of spectrotemporal analysis in mental health assessment.