A cross-linguistic depression detection method based on speech data.

Background: Depression is a common and disabling psychological disorder that affects patients and their social circles. With advances in artificial intelligence, researchers have focused on automated depression detection using speech data. However, most current AI methods rely on monolingual data, limiting their cross-linguistic generalization.

Methods: We propose a transfer learning method called Deep Covariance Alignment Network (DCAN) that transfers models trained on English speech data (source domain) to Chinese speech data (target domain). Experiments were conducted on the DAIC-WOZ and MODMA datasets. We used down-sampled speech data (1 kHz), features extracted by a Convolutional AutoEncoder, and manually selected features to explore commonalities across languages and compare our method with other models.

Results: Our model achieved an accuracy of 88.7 % on the English dataset and 81.1 % on the Chinese dataset, outperforming models trained solely on English data by an average of 21.9 %. Compared to other mainstream transfer learning methods, our approach showed a 4 % improvement. This reveals that even across different linguistic and cultural backgrounds, there are potential commonalities between speech signals and depression.

Limitations: Future research should incorporate a wider range of languages.

Conclusions: These findings highlight that our research enhances the generalization capability of depression detection models based on speech data across different linguistic domains, thereby reducing the effort required for constructing diverse language-specific speech datasets. The method we proposed are expected to be useful in supporting the diagnosis of depressions.