ALSA-CNN-transformer: audio depression detection using hybrid attention and multi-scale feature fusion

Depression is a widespread and debilitating mental disorder. Conventional diagnostic methods relying on clinical interviews and the Self-Rating Depression Scale suffer from subjective bias, low screening efficiency, and poor performance in early intervention. To overcome the drawbacks of inadequate feature extraction and weak sequence modeling in existing audio depression detection algorithms, we propose a novel model for Chinese speech-based depression screening. This paper proposes a Six-layer Alternating Local-Sparse-Atrous Attention Convolutional Neural Network-Transformer (ALSA-CNN-Transformer) model. After standardized audio preprocessing, the model extracts Mel-spectrogram features via sixth-order complex Gaussian Continuous Wavelet Transform (cgau6 CWT) and dual-band Chroma features, then fuses the two feature branches into joint representations for capturing spectral and pitch abnormalities of depressive speech. One-dimensional convolutional layers extract local fine-grained spectral features, which are subsequently fed into a six-layer Transformer encoder equipped with alternating local, sparse, and atrous self-attention to simultaneously capture short-term local speech details and long-range emotional temporal dependencies. Finally, adaptive pooling and fully connected layers are adopted to complete binary depression classification. On the Chinese EATD-Corpus, the model achieves a precision of 0.9531, a recall of 0.9524, and an F1-score of 0.9525. It shows strong robustness to noise and supports remote and large-scale screening. This efficient, objective, noninvasive approach assists depression diagnosis, promotes early detection, and helps reduce mental health stigma.