Ensemble learning with speaker embeddings in multiple speech task stimuli for depression detection.

Area of Science:

• Speech Signal Processing
• Machine Learning for Healthcare
• Mental Health Diagnostics

Background:

• Speech signals are promising, non-invasive biomarkers for depression detection.
• Challenges include speech variability, limited depression speech data, and feature length.
• Existing methods struggle with diverse acoustic and emotional speech variations.

Purpose of the Study:

• To propose a multi-task ensemble learning method for robust depression classification from speech.
• To address limitations of data scarcity and speech variability using speaker embeddings.
• To enhance depression detection accuracy by fusing information from diverse speech tasks.

Main Methods:

• Extracted Mel Frequency Cepstral Coefficients (MFCC), Perceptual Linear Predictive Coefficients (PLP), and Filter Bank (FBANK) features.
• Trained speaker embedding extractors (Resnet x-vector, TDNN x-vector, i-vector) on an out-domain dataset.
• Employed multi-task learning with SVM and RF, followed by MLP for final classification, fusing results from nine speech tasks.

Main Results:

• MFCC-based Resnet x-vectors demonstrated superior performance in depression detection.
• Interview speech and neutral emotional stimuli yielded better recognition results.
• The proposed multi-task ensemble learning method, particularly combining MFCC and PLP features, significantly improved depression identification accuracy.

Conclusions:

• The multi-task ensemble learning approach effectively fuses depression-related information from various stimuli.
• This method offers a promising new avenue for non-invasive depression detection using speech.
• Future work includes pre-training on augmented in-domain datasets to further enhance performance.