FRAUG: A FRAME RATE BASED DATA AUGMENTATION METHOD FOR DEPRESSION DETECTION FROM SPEECH SIGNALS.

Area of Science:

• Computational linguistics
• Speech processing
• Machine learning for healthcare

Background:

• Depression detection from speech is a growing area of research.
• Existing data augmentation methods often alter acoustic properties directly.
• A need exists for data augmentation techniques that modify feature representations without explicit acoustic manipulation.

Purpose of the Study:

• To propose a novel data augmentation method for speech-based depression detection.
• To evaluate the effectiveness of the proposed method against baseline and established augmentation techniques.
• To demonstrate the generalizability of the method across different datasets, models, and acoustic features.

Main Methods:

• Data augmentation by modifying frame-width and frame-shift during feature extraction.
• Altering time-frequency resolution of frame-level features, not explicit acoustic parameters.
• Evaluation on DAIC-WOZ (English) and CONVERGE (Mandarin) datasets using DepAudioNet, CNN, mel-Spectrograms, x-vector embeddings, and MFCCs.

Main Results:

• Significant improvements in depression detection accuracy: 5.97% (validation) and 25.13% (test) on DAIC-WOZ.
• Improvements of 9.32% (validation) and 12.99% (test) on CONVERGE.
• Outperformed baseline systems and common augmentation methods like noise, VTLP, Speed, and Pitch Perturbation.

Conclusions:

• The proposed data augmentation method effectively enhances depression detection from speech signals.
• Modifying time-frequency resolution offers a distinct advantage over traditional acoustic manipulation augmentation.
• The method shows robust performance across diverse datasets and model architectures, indicating its practical utility.