Depression recognition using a proposed speech chain model fusing speech production and perception features.

Area of Science:

• Computational linguistics
• Speech processing
• Mental health informatics

Background:

• Increasing depression diagnoses strain clinical resources.
• Audio-based screening is a valuable tool for early detection.
• Current methods overlook vocal tract changes, impacting recognition accuracy.

Purpose of the Study:

• To propose a novel machine speech chain model for depression recognition (MSCDR).
• To improve depression recognition by integrating speech production and perception features.
• To capture text-independent depressive speech representations.

Main Methods:

• Utilizing linear predictive coding (LPC) for speech generation and Mel-frequency cepstral coefficients (MFCC) for speech perception.
• Employing a one-dimensional convolutional neural network (CNN) and a long short-term memory (LSTM) network.
• Sequentially capturing dynamic features for classification.

Main Results:

• Achieved accuracies of 0.77 and 0.86 on two diverse datasets.
• Obtained average F1 scores of 0.75 and 0.86, outperforming existing methods.
• Demonstrated the complementary value of speech production and perception features.

Conclusions:

• The proposed MSCDR shows strong generalization and superiority.
• Vocal tract changes in depression warrant attention for audio-based diagnosis.
• The model offers a promising auxiliary tool for depression screening.