Depression Classification Using Log-Mel Spectrograms: A Comparative Analysis of Window Size-Based Data Augmentation and Deep Learning Models

In this paper, we presents an innovative approach to detecting depression by analyzing log-mel spectrograms from speech recordings of depressed and non-depressed speakers. As an augmentation strategy, we leverage multiple window sizes for spectrogram generation to address class imbalance and explore various deep learning models for binary classification. U sing the Emotion Audio-Textual Depression corpus(EATD-Corpus), a Chinese language depression dataset, four models, VGG-16, ResNet50, EfficientNet, and ConvMixer, were trained and compared. Results show that multiple window sizes significantly enhance classification accuracy, with ConvMixer performing exceptionally well on multi-resolution log-mel spectrograms. Smaller windows capture finer temporal details and high-frequency content, while larger windows improve frequency resolution and low-frequency information. For the ConvMixer model, we achieved an accuracy of 80.90% and F-l scores of 0.79 and 0.82 for depressed and non-depressed classes, respectively. This research offers valuable insights into mental health diagnostics and could aid professionals in more efficient diagnostics.