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

TL;DR

This work addresses semantic bias in speech-based depression detection caused by entanglement between linguistic sentiment and diagnostic labels. It introduces DepFlow, a three-stage framework that disentangles depressive acoustic cues via a Depression Acoustic Encoder, generates waveform-level, depression-conditioned speech with a flow-based TTS (DepFlow), and employs a prototype-based severity mapper for smooth, interpretable control over depressive expressiveness. By constructing Camouflage Depression-oriented Augmentation (CDoA) that pairs depressive acoustics with positive/neutral content, the approach improves macro-F1 and robustness across several detection architectures, outperforming conventional augmentation strategies. The framework provides a controllable data-generation platform for robustness testing and simulation in mental-health speech systems, while acknowledging ethical considerations and the need for broader validation across populations and languages.

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.

Figures (6)

• Figure 1: Training pipeline of DepFlow. DepFlow takes phoneme sequences, speaker embeddings, and a depression condition embedding $\mathbf{c}_{\mathrm{dep}}$ as conditioning inputs. A text encoder provides linguistic features and a duration module aligns them with acoustic frames, while a Conditional Flow Matching decoder generates mel-spectrograms with FiLM-based depression conditioning.
• 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 $\mathbf{d}$, with an ordinal regression head estimating PHQ-based severity and GRL-based speaker and content disentanglement heads suppressing speaker identity and linguistic information.
• Figure 3: Inference pipeline of DepFlow. Given a desired severity level, DepFlow maps it to a depression condition embedding $\mathbf{c}_{\mathrm{dep}}$ and, together with phoneme and speaker inputs, generates severity-controlled mel-spectrograms via FiLM-conditioned decoding, which are finally converted to speech using HiFi-GAN.
• Figure 4: 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. Red tiles indicate overrepresentation (observed $>$ expected), while blue tiles indicate underrepresentation. The deep red shading in the Depressed-Negative intersection highlights a significant correlation between depression labels and negative linguistic sentiment.
• Figure 5: UMAP visualization of the DAE embeddings. (a) Embeddings obtained from a frozen WavLM-Large model; (b) Embeddings of the same utterances after processing through the trained DAE. In each panel, the five subplots correspond respectively to the five clinical severity levels (Healthy, Mild, Moderate, Moderately Severe, Severe): red points denote samples from the target severity level, while blue points denote all remaining samples.