DepFlow: Disentangled Speech Generation to Mitigate Semantic Bias in Depression Detection

Abstract: Speech is a scalable and non-invasive biomarker for early mental health screening. However, widely used depression datasets like DAIC-WOZ exhibit strong coupling between linguistic sentiment and diagnostic labels, encouraging models to learn semantic shortcuts. As a result, model robustness may be compromised in real-world scenarios, such as Camouflaged Depression, where individuals maintain socially positive or neutral language despite underlying depressive states. To mitigate this semantic bias, we propose DepFlow, a three-stage depression-conditioned text-to-speech framework. First, a Depression Acoustic Encoder learns speaker- and content-invariant depression embeddings through adversarial training, achieving effective disentanglement while preserving depression discriminability (ROC-AUC: 0.693). Second, a flow-matching TTS model with FiLM modulation injects these embeddings into synthesis, enabling control over depressive severity while preserving content and speaker identity. Third, a prototype-based severity mapping mechanism provides smooth and interpretable manipulation across the depression continuum. Using DepFlow, we construct a Camouflage Depression-oriented Augmentation (CDoA) dataset that pairs depressed acoustic patterns with positive/neutral content from a sentiment-stratified text bank, creating acoustic-semantic mismatches underrepresented in natural data. Evaluated across three depression detection architectures, CDoA improves macro-F1 by 9%, 12%, and 5%, respectively, consistently outperforming conventional augmentation strategies in depression Detection. Beyond enhancing robustness, DepFlow provides a controllable synthesis platform for conversational systems and simulation-based evaluation, where real clinical data remains limited by ethical and coverage constraints.

• The paper introduces a three-stage framework (DepFlow) that disentangles depressive acoustic cues from semantic content to address bias in depression detection.
• The methodology employs a Depression Acoustic Encoder, a flow-matching TTS model, and prototype-based severity mapping to control depressive expressiveness.
• Evaluation shows significant improvements in macro-F1 scores, demonstrating the framework's potential to enhance the robustness of depression detection systems.

DepFlow: Disentangled Speech Generation to Mitigate Semantic Bias in Depression Detection

Introduction

The research paper "DepFlow: Disentangled Speech Generation to Mitigate Semantic Bias in Depression Detection" (2601.00303) explores a novel framework called DepFlow, designed to address semantic bias in datasets used for depression detection via speech analysis. Semantic bias, particularly the coupling between linguistic sentiment and depression labels, has been identified as a critical challenge in the robustness of depression detection models. This paper proposes a three-stage framework, DepFlow, to mitigate this bias by enabling the disentangled generation of speech samples where depression-related acoustic cues can be independently controlled.

Methodology

DepFlow is composed of three key components: the Depression Acoustic Encoder (DAE), a flow-matching TTS model, and a prototype-based severity mapping mechanism. The DAE is tasked with learning speaker- and content-invariant depression embeddings through adversarial training, ensuring that depressive cues remain discernible while disentangling speaker identity and linguistic content. The flow-matching TTS model utilizes these embeddings to control depressive severity in synthesized speech samples, maintaining the original content and speaker characteristics. Finally, the prototype-based severity mapping allows for smooth and interpretable manipulation of depressive expressiveness across a defined continuum.

Figure 1: Training pipeline of DepFlow. DepFlow takes phoneme sequences, speaker embeddings, and a depression condition embedding $\mathbf{c}_{\mathrm{dep}}$ as conditioning inputs.

In the pursuit of minimizing semantic bias, the research introduces a Camouflage Depression-oriented Augmentation (CDoA) dataset, which combines depressive acoustic patterns with positive or neutral semantic content. This novel data augmentation strives to decouple sentiment from depressive cues, creating underrepresented scenarios in conventional datasets.

Evaluation

The effectiveness of DepFlow is demonstrated by evaluating its performance across multiple depression detection architectures. The introduction of the CDoA dataset resulted in significant macro-F1 improvements of 9%, 12%, and 5% respectively. These results suggest substantial gains in the robustness and accuracy of depression detectors when augmented with DepFlow-generated data.

Figure 2: Architecture of the Depression Acoustic Encoder (DAE). Frame-level WavLM features are aggregated into an utterance-level representation to produce a depression acoustic embedding $\mathbf{d}.$

Moreover, the study performs an in-depth analysis of the DAIC-WOZ dataset to understand the extent of semantic bias, as evidenced by a strong coupling between negative sentiments and depression labels (Figure 3). The ability of DepFlow to neutralize this bias indicates its potential as a powerful tool for building more resilient depression detection systems.

Figure 3: Analysis of semantic bias in the DAIC-WOZ dataset. (a) Sentiment Distribution by Diagnosis Groups; (b) Mosaic plot of Diagnosis vs. Sentiment shaded by Pearson residuals.

Implications and Future Work

The implications of DepFlow extend beyond its immediate use in depression detection. By allowing the synthesis of speech with controllable depression-related acoustic characteristics, the framework offers a principled approach to data generation for dialogue-based studies, conversational agents, and simulation-based evaluations where real-world data is constrained by ethical and practical limitations.

The research hints at promising directions for future work, including validation against broader and more varied datasets, extending language capabilities, and enhancing ethical safeguards to balance the benefits of synthetic data generation with its potential misuses. These advances are crucial for deploying AI in sensitive contexts such as mental health, where reliability and ethical considerations intersect.

Conclusion

DepFlow presents a compelling approach to generate speech data that disentangles semantic content from depression-related acoustic cues. Its ability to systematically address semantic bias in depression detection datasets marks an important stride towards more robust and clinically relevant AI applications in mental health assessment. Future developments of DepFlow could significantly advance the field by providing enhanced tools for the simulation, evaluation, and modeling of mental health conditions.