Related Experiment Video
Updated: Oct 7, 2025
06:37
Author Spotlight: Addressing Technical and Subjective Challenges in Measuring Classroom Attention
Published on: December 15, 2023
4.3K
Speech depression recognition based on attentional residual network.
Xiaoyong Lu1,2, Daimin Shi3, Yang Liu3
1Internet Education Data Learning Analysis Technology National and Local Joint Engineering Laboratory, 730070 Lanzhou, Gansu, China.
Frontiers in Bioscience (Landmark Edition)
|January 7, 2022
Summary
This study introduces an attention residual network to accurately predict depression from speech. The model outperforms traditional methods, offering a practical tool for depression recognition and severity assessment.
Area of Science:
- Speech analysis
- Computational psychiatry
- Machine learning
Background:
- Depressive disorder, characterized by sadness and loss of interest, impacts concentration and self-worth.
- Speech analysis offers a non-offensive, low-cost method for depression prediction.
- Existing methods struggle with complex deep learning structures and manual feature extraction.
Purpose of the Study:
- To develop an advanced speech depression recognition model.
- To address limitations of current deep neural networks and traditional machine learning approaches.
- To improve the accuracy and practicality of depression detection using speech signals.
Main Methods:
- A novel attention residual network model was designed.
- A depression corpus was created using the self-reference effect (SRE) paradigm.
- Channel and spatial attention mechanisms were integrated into residual units for feature learning.
Main Results:
- The proposed model significantly outperformed traditional machine learning methods in depression recognition.
- The model demonstrated practical applicability for real-world depression recognition needs.
- Spontaneous speech yielded better results than automatic speech.
Conclusions:
- The study successfully predicted depression and estimated its severity using speech.
- Speech features from negative questions and negative emotions showed better classification accuracy.
- Higher recognition accuracy was observed for both male and female subjects across various emotional contexts.
