disentanglement using pre-trained features

This article proposes using pre trained features to address the challenge of detecting depression through speech. Traditional raw audio has shown low accuracy and insufficient generalization performance in depression detection. We use pre trained models that have been developed to extract features, which can be used to extract general feature representations from speech data. During the pre training process, we further decouple the speakers, introducing prior information and providing a better starting point for training downstream models. The results indicate that we achieved the best performance when using the extracted features from the CONTANTVEC pre- trained model with speaker decoupling improvement.