A Privacy-Preserving Unsupervised Speaker Disentanglement Method for Depression Detection from Speech

Area of Science:

• Computational linguistics
• Machine learning
• Speech processing

Background:

• Current depression detection from speech often requires speaker labels, leading to privacy concerns.
• Existing methods can be unstable and add complexity through adversarial domain prediction.

Purpose of the Study:

• To develop an unsupervised speaker disentanglement method for privacy-preserving depression detection.
• To improve the accuracy and reduce the complexity of speech-based depression detection systems.

Main Methods:

• An unsupervised approach reducing cosine similarity between latent spaces of depression and speaker classification models.
• Utilizing ComparE16 features and an LSTM-only model on the DAIC-WOZ dataset.
• Score-level fusion with a Word2vec-based text approach.

Main Results:

• Achieved an F1-Score of 0.776 and a speaker de-identification (DeID) score of 92.87%, outperforming adversarial methods.
• Demonstrated superior performance compared to an adversarial counterpart (F1-Score 0.762, DeID 68.37%).
• Score-level fusion enhanced performance to an F1-Score of 0.830.

Conclusions:

• The proposed unsupervised method enhances privacy in speech-based depression detection without speaker labels.
• This approach reduces model complexity and improves performance over baseline and adversarial methods.
• Speaker disentanglement is complementary to text-based methods, offering significant performance gains when combined.