CDoA: Camouflage Depression Augmentation Dataset

• CDoA is a specialized speech dataset that challenges depression detectors by pairing benign text with depressive acoustic features to expose camouflaged diagnostic patterns.
• It employs a depression-conditioned TTS framework with adversarial embedding disentanglement and interpretable PHQ-8 severity control to produce clinically meaningful synthetic utterances.
• Evaluations on multiple models demonstrate that CDoA improves depression detection performance by up to 12 points, reducing reliance on sentiment cues.

The Camouflage Depression-oriented Augmentation (CDoA) dataset is a speech augmentation resource specifically designed to address confounded sentiment–diagnosis correlations and semantic shortcut vulnerabilities in depression detection models. Constructed via a depression-conditioned neural text-to-speech (TTS) framework, CDoA contains synthetic utterances that systematically pair positive or neutral lexical content with acoustics characteristic of clinical depression, thereby exposing detectors to “camouflaged” patterns where sentiment is decorrelated from depressive state. The dataset is anchored in the DAIC-WOZ corpus, leverages advanced adversarial embedding disentanglement, and enables fine-grained severity control of depressive acoustics/phonation. CDoA demonstrably improves downstream detection robustness relative to conventional augmentation strategies and is available for research use under a CC-BY-NC license (Li et al., 1 Jan 2026).

1. Motivation and Problem Setting

State-of-the-art depression detection systems trained on datasets such as DAIC-WOZ exhibit a clinically spurious coupling between lexical sentiment (e.g., negative word usage) and diagnostic label. This coupling encourages models to learn “semantic shortcuts,” wherein detectors anchor predictions on the sentiment polarity of input words rather than on speech-invariant, clinically robust acoustic markers. In real-world scenarios, “Camouflaged Depression” frequently arises when individuals intentionally or subconsciously present positive or neutral content while retaining vocal signatures consistent with depressive conditions—such as psychomotor retardation, low energy, or breathiness.

The CDoA resource was engineered to break the sentiment–diagnosis confound by constructing speech examples where benign (positive/neutral) text is rendered with the acoustic properties of depression. This generates acoustic-semantic mismatches underrepresented in natural corpora, directly challenging detectors to shift their reliance from text sentiment toward depression-specific biomarkers.

2. Data Sources and Sentiment-Stratified Text Processing

All lexical content in CDoA originates from the official DAIC-WOZ training transcripts. The data preprocessing pipeline involves diarization, silence filtering, and the selection of utterances longer than one second, yielding 9,234 interviewee samples. Sentiment annotation is performed using the DeepSeek-R1 LLM, stratifying the corpus with approximately one-third of utterances mapped to each sentiment label (positive, neutral, negative).

The text prompts are grouped as follows:

• Positive/neutral utterances form the “benign text” bank.
• Negative utterances constitute the “depressive text” bank.

This two-bank design maximizes control over the semantic polarity injected into each synthetic waveform, allowing CDoA to create controlled acoustic-semantic pairings for augmentation.

3. Depression-Specific Acoustic Embedding and Controllable Synthesis

To extract depression-specific prosodic features disentangled from speaker identity and phonetic content, a Depression Acoustic Encoder (DAE) is trained on DAIC-WOZ audio:

• Feature extraction: Frame-level WavLM-Large representations ($x_t \in \mathbb{R}^{768}$) are projected to 256-dim vectors, aggregated via an attention-based pooling mechanism, then mapped to a 32-dim depression embedding ($d$).
• Multi-head decomposition:
  • Ordinal regression head estimates PHQ-8 severity levels using monotonic binary cross-entropy.
  • Speaker classification (with adversarial Gradient Reversal Layer) removes speaker identity.
  • Content adversarial head discourages retention of linguistic content.
• Composite loss: $$L_{total} = \lambda_{sup}L_{sup} + \lambda_{id}L_{id} + \lambda_{spk}L_{adv-spk} + \lambda_{con}L_{adv-con}$$, with $$\{\lambda_{sup},\lambda_{id},\lambda_{spk},\lambda_{con}\} = \{1.0, 0.2, 0.2, 0.1\}$$.

The trained DAE achieves utterance-level ROC-AUC of 0.693 for depression severity, exhibits high speaker verification EER (>0.35), and low content reconstruction, confirming effective disentanglement.

The TTS backbone, DepFlow, is a Matcha-TTS variant that accepts phoneme sequences, fixed speaker embeddings, and a continuous depression embedding as conditions. Depression control is effected via FiLM modulation: a small MLP maps the 32-dim control vector into FiLM scale–shift parameters across all decoder blocks, permeating depressive cues into the generated acoustic features without compromising lexical or speaker fidelity.

For interpretable severity control, subject-level depression embeddings are aggregated by PHQ-8 bin to create prototype vectors. At inference, desired PHQ scores $s\in[0,24]$ are normalized, interpolated between adjacent prototypes via spherical linear interpolation (SLERP), and injected as the depression control vector—enabling continuous, smooth variation across the depression spectrum.

4. Construction Pipeline and Dataset Characteristics

CDoA is synthesized via the following steps:

1. Sentiment annotation: DeepSeek-R1 inference on all DAIC-WOZ training utterances, forming benign vs. depressive text banks.
2. Synthesis quotas: Balanced quotas per subject across PHQ bins are enforced: 13 (healthy), 34 (mild), 2 (moderate), 91 (moderately severe), and 194 (severe) samples per individual.
3. Sample synthesis workflow:
   • For targets labeled depressed, randomly select a benign text prompt (to induce camouflage mismatch).
   • For targets labeled healthy, randomly select a depressive text prompt.
   • Convert to phonemes, retrieve speaker embedding, compute depression control vector using SLERP at the subject’s PHQ score.
   • Synthesize waveform via DepFlow (FiLM-conditioning + HiFi-GAN vocoder); each sample is approximately 10 seconds at 22.05 kHz.
   • Annotate metadata: speaker_id, text, sentiment, PHQ score, binary depression label, condition (“synthetic” or “natural”).
4. Dataset composition: 5,760 synthetic utterances combined with original DAIC training utterances, yielding a final corpus of 5,760 training examples (2,880 depressed, 2,880 non-depressed).

Basic statistics and metadata fields are summarized below:

Field	Description/Distribution
Num. synthetic samples	5,760
Speakers	107 DAIC-WOZ subjects (mixed gender)
Per-subject quotas	{13, 34, 2, 91, 194} by PHQ-8 bin
Duration	≈10 s/sample, 22.05 kHz, 16-bit PCM mono WAV
Metadata	speaker_id, text, sentiment, PHQ score, binary_label, condition

5. Quantitative Evaluation and Comparative Analysis

The impact of CDoA was benchmarked on three depression detection architectures:

• DepAudioNet: CNN-RNN on log-mel spectrograms
• NUSD: ECAPA-TDNN with speaker-disentangling branches
• HAREN-CTC: WavLM-based waveform-level model

Following the DAIC-WOZ partition (107/35/47 subjects for train/dev/test), synthetic CDoA examples were only included in training. Each epoch involved 10-second crops; test inference applied majority voting over each subject’s 20 longest utterances. Performance was measured via Macro-F1, Sensitivity, and Specificity at the subject level.

CDoA yielded consistent improvements over unaugmented baselines:

Model	Macro-F1 (Baseline)	Macro-F1 (+CDoA)	Relative Gain
DepAudioNet	0.482	0.526	+9 points
NUSD	0.514	0.577	+12 points
HAREN-CTC	0.525	0.551	+5 points

CDoA also surpassed SpecAugment, FrAUG, Mixup, and generic “Instruct-TTS” augmentation strategies across all metrics.

6. Usage, Ethics, and Release

CDoA is intended for integration by simple concatenation to training sets, maintaining class balance through explicit binary and PHQ labeling. Each synthetic example is fully annotated for provenance, facilitating controlled studies and ablation analyses.

The generation of synthetic depressive speech imposes significant ethical obligations, especially in clinical contexts. Recommended safeguards include inclusion of provenance metadata, audio watermarking (e.g., inaudible tags), and institutional data-use agreements. Informed consent and transparent reporting are advised when deploying models trained in part on CDoA-augmented data.

All audio files, metadata, models (DepFlow), and construction recipes are released under a CC-BY-NC license and are accessible at https://github.com/ntu-spmlab/DepFlow or by direct author request (Li et al., 1 Jan 2026).

7. Research Implications and Future Directions

CDoA provides a focused mechanism for addressing critical limitations in current depression detection benchmarks—specifically, the tendency of models to overfit to semantic sentiment as a diagnostic proxy. By exposing models to camouflaged patterns, CDoA compels a reweighting toward clinically grounded acoustic biomarkers, enhancing robustness to “realistic” scenarios in which individuals with depression mask their symptoms lexically.

A plausible implication is that CDoA’s construction protocol, including interpretable depression manifolds and adversarial disentanglement, can be adapted for related clinical speech applications with similar confounding structures. The platform is positioned as a reusable tool for simulation-based evaluation in affective computing and speech-based mental health assessment, particularly where ethical and coverage constraints limit collection of real clinical data.