DAIC-WOZ: Multimodal Depression Interview Corpus

• DAIC-WOZ is a multimodal corpus of semi-structured clinical interviews that integrates transcripts, audio, and visual data with PHQ-8-linked depression annotations.
• It employs a Wizard-of-Oz setup where a human-controlled virtual interviewer, Ellie, systematically guides the interview, enabling research on depression detection and methodological biases.
• Benchmark studies using DAIC-WOZ reveal challenges in reproducibility, shortcut learning from interviewer prompts, and demographic fairness due to modality and label inconsistencies.

The Distress Analysis Interview Corpus/Wizard-of-Oz (DAIC-WOZ) is a multimodal corpus of semi-structured clinical interviews collected through interactions between a human participant and a virtual interviewer named Ellie, whose behavior is controlled by a human operator in a Wizard-of-Oz setup. It has become a central benchmark for automatic depression analysis because it combines interview transcripts, audio, video, and derived feature streams with PHQ-8-linked depression annotations, and because subsets of the corpus were used in AVEC 2016, 2017, and 2019 challenge settings (Tasnim et al., 2023, Qureshi et al., 2019). At the same time, DAIC-WOZ has increasingly been treated not only as a benchmark but also as a methodological case study: recent work argues that the corpus’s semi-structured interviewer behavior can itself leak label-related information, so strong performance may reflect shortcut learning from Ellie’s prompts rather than modeling participant language or behavior (Burdisso et al., 2024, Watawana et al., 25 Mar 2026).

1. Definition, interview protocol, and corpus design

DAIC-WOZ is described as a corpus of semi-structured clinical interviews in English designed to support diagnosis or analysis of depression, anxiety, and post-traumatic stress disorder. The interviews are conducted by Ellie, an animated virtual interviewer, but the interaction is not fully autonomous: Ellie is human-controlled, which is the defining Wizard-of-Oz property emphasized in work that analyzes the corpus’s methodological assumptions (Burdisso et al., 2024, Tasnim et al., 2023).

The corpus is repeatedly characterized as multimodal. Across the cited studies, the released material includes transcripts, raw interview audio, and derived visual and acoustic streams such as 68-point facial landmarks, head pose, eye gaze, facial action units from CERT, COVAREP features, and formant features (Qureshi et al., 2019). Other studies summarize DAIC-WOZ more compactly as containing transcripts, audio, video, and questionnaire-linked information for 189 interview sessions (Lorenzoni et al., 2024).

A central design feature is the semi-structured interview script. Ellie is described as having a finite repertoire of 191 prompts, including general questions, neutral backchannels, positive empathy, negative empathy, surprise responses, continuation prompts, and miscellaneous prompts (Burdisso et al., 2024, Watawana et al., 25 Mar 2026). This protocol was intended to provide consistency, but later analyses argue that the same consistency creates predictable regularities in prompt identity, prompt ordering, and prompt position. In that framing, DAIC-WOZ is not simply a resource for depression modeling; it is also a resource for studying what semi-structured interaction protocols make available to machine learning systems (Watawana et al., 25 Mar 2026).

2. Data organization, labels, and task formulations

Several papers report that DAIC-WOZ contains 189 participants or interview sessions, with official train, development or evaluation, and test partitions of 107, 35, and 47 interviews respectively (Watawana et al., 25 Mar 2026, Bailey et al., 2020). Many studies report development or evaluation results because the official test labels were not public under the AVEC protocol (Burdisso et al., 2024, Bailey et al., 2020).

The principal supervisory signal is PHQ-8. Some papers use the PHQ-8 total score directly as a regression target; others derive binary depressed/not-depressed labels; still others define five PHQ-8-based severity classes. The operational details are not fully uniform across the literature. For example, one audio-bias study defines non-depressed as $r<10$ and depressed as $r\ge10$ (Bailey et al., 2020), whereas another severity-classification paper uses five PHQ-8-based categories and also describes a binary split of PHQ-8 $\le 10$ versus $>10$ for some analyses (Mao et al., 2022). This suggests that DAIC-WOZ supports several task formulations, but that cross-paper comparisons require close attention to label definitions.

Formulation	Target	Representative uses
Binary depression classification	PHQ-8-derived depressed/control label	(Burdisso et al., 2024, Watawana et al., 25 Mar 2026, Lorenzoni et al., 2024, Junias et al., 30 Sep 2025, Bailey et al., 2020)
Depression severity regression	PHQ-8 total score in $[0,24]$	(Qureshi et al., 2019)
Five-class severity classification	Healthy, mild, moderate, moderately severe, severe	(Mao et al., 2022)

At the participant level, DAIC-WOZ has also been used for subgroup analysis. One fairness study reports 186 participants after excluding 3 with missing gender and 184 after excluding 5 with missing race. It gives gender-group counts of female: 23 depression / 62 no depression and male: 19 depression / 82 no depression, and race/ethnicity counts of African American: 9/51, Hispanic: 3/28, and White/Caucasian: 21/78 (Junias et al., 30 Sep 2025). These counts are frequently cited because they quantify the small subgroup sizes that later fairness analyses identify as a major limitation.

3. Modalities and benchmark use in modeling research

DAIC-WOZ has been used through text-only, audio-only, visual-only, and multimodal pipelines. In multimodal PHQ-8 regression, one AVEC-oriented study predicted the eight PHQ-8 item scores separately and summed them to obtain a total in $[0,24]$, using participant-isolated audio, landmark-based visual features, and compact transcript-derived lexical features, followed by decision-level fusion (Dham et al., 2017). Another multimodal study, framed around interview-level PHQ-8 regression, used seven streams grouped into text, acoustic, and visual modalities: transcript utterances encoded with the Universal Sentence Encoder, COVAREP and formants, and visual streams comprising facial landmarks, head pose, eye gaze, and action units (Qureshi et al., 2019).

DAIC-WOZ has also supported sequence-based severity classification. One study converted each interview into many fixed-length audio and text segments, performed sequence-level prediction, and aggregated to a patient-level label by majority voting. For audio it used DAIC-WOZ’s precomputed COVAREP features every 10 ms, removed frames with VUV $=0$, and retained 73-dimensional voiced frame vectors; for text it removed stop words with NLTK, applied WordNet lemmatization, and embedded tokens with non-trainable 100-dimensional GloVe vectors (Mao et al., 2022).

Text-only use cases span both classical and neural modeling. A transcript-only case study constructed hand-engineered tabular features from answer-level sentiment, response timing, speech speed, lexical frequencies, and part-of-speech statistics, then compared Random Forest, XGBoost, and SVM models on a custom 80/20 split rather than the official DAIC-WOZ partitions (Lorenzoni et al., 2024). More recent work uses long-document transformers, graph convolutional networks over interview-word graphs, Mental-RoBERTa with chunking of question-answer pairs into overlapping 300-word windows, and few-shot prompting of GPT-4o-mini on full clinical interview transcripts (Watawana et al., 25 Mar 2026, Junias et al., 30 Sep 2025).

This breadth of modeling practice has made DAIC-WOZ useful for benchmarking, but it also means that “DAIC-WOZ results” aggregate heterogeneous choices about modality inclusion, speaker filtering, segmentation, label definition, and evaluation protocol. That heterogeneity is one reason later papers treat DAIC-WOZ as a benchmark that requires methodological scrutiny rather than simple score comparison.

4. Interviewer prompts as context and as confound

The most consequential recent controversy concerns the use of Ellie’s prompts. Earlier work often treated interviewer text as contextual information, but two dedicated analyses argue that interviewer turns in DAIC-WOZ can function as a shortcut signal because the semi-structured protocol makes some prompts, prompt combinations, and prompt positions label-correlated (Burdisso et al., 2024, Watawana et al., 25 Mar 2026).

The clearest diagnostic is a speaker ablation: participant-only versus interviewer-only text. In one focused critique on the DAIC-WOZ evaluation split, participant-only LongBERT reached macro $F_1=0.72$ and participant-only GCN reached $0.85$, while interviewer-only LongBERT reached $0.84$ and interviewer-only GCN reached $r\ge10$0. The same paper then combined participant-only and interviewer-only GCN predictions with a conjunctive rule and obtained macro $r\ge10$1, which it explicitly presented as intentional exploitation of the bias (Burdisso et al., 2024). A later cross-dataset study using official DAIC-WOZ splits reported the same qualitative pattern on development: P-Longformer $r\ge10$2 versus I-Longformer $r\ge10$3, and P-GCN $r\ge10$4 versus I-GCN $r\ge10$5. On the held-out test split, interviewer-only still exceeded participant-only for GCN, $r\ge10$6 versus $r\ge10$7, though not for Longformer, where participant-only was higher, $r\ge10$8 versus $r\ge10$9 (Watawana et al., 25 Mar 2026).

The mechanism proposed in these studies is highly specific. Interviewer-only models concentrate on a very small subset of prompts, especially “How do you cope with that?”, “Do you still go to therapy?”, and “Do you feel therapy is useful?”, while largely ignoring other clinically relevant questions (Watawana et al., 25 Mar 2026). A first-half versus second-half ablation sharpens the point: in one study, E-longBERT scored $\le 10$0 on the full interview, $\le 10$1 on the first half only, and $\le 10$2 on the second half only, whereas participant-only performance was much more evenly distributed across interview halves (Burdisso et al., 2024).

Heatmap analyses localize the same effect in time. Interviewer-only GCN evidence forms narrow, high-contrast bands at specific normalized positions, while participant-only evidence is distributed more broadly across the conversation (Watawana et al., 25 Mar 2026). The papers interpret this as evidence that interviewer-side models are exploiting fixed prompt locations within the semi-structured script. The practical implication is not that interviewer context is useless in principle; rather, in DAIC-WOZ’s current format interviewer turns are treated as dangerously confounded with the label unless a study can demonstrate otherwise (Burdisso et al., 2024, Watawana et al., 25 Mar 2026).

5. Fairness, subgroup imbalance, and demographic bias

A second major line of criticism concerns demographic bias. One audio study examined DAIC-WOZ’s training split and reported 44 female participants and 63 male participants, with female counts of 27 non-depressed and 17 depressed and male counts of 49 non-depressed and 14 depressed. It summarized this as $\le 10$3, arguing that acoustic models can exploit gender-correlated cues because depression prevalence differs by gender within the released split (Bailey et al., 2020).

That study rebalanced the four gender-class quadrants $\le 10$4, $\le 10$5, $\le 10$6, and $\le 10$7 down to the smallest cell size of 14 and retrained mel-spectrogram and raw-audio models. The reproduced DepAudioNet variants then dropped from $\le 10$8 and $\le 10$9 average $>10$0 to $>10$1 and $>10$2, respectively, whereas raw-audio variants were less affected and in some settings improved after balancing (Bailey et al., 2020). The authors interpreted the large post-balancing performance loss as evidence that uncorrected DAIC-WOZ experiments can overreport depression-detection performance.

A more recent transcript-only fairness study broadened the analysis from gender to race/ethnicity and compared a chunked Mental-RoBERTa baseline with GPT-4o-mini few-shot prompting (Junias et al., 30 Sep 2025). Before debiasing, the DNN baseline had male balanced accuracy $>10$3 and female balanced accuracy $>10$4, with equalized-odds score $>10$5. Race disparities were sharper: Hispanic performance was degenerate at TPR $>10$6, precision $>10$7, $>10$8, BA $>10$9, while White/Caucasian BA was $[0,24]$0 (Junias et al., 30 Sep 2025).

The same study found that worst-group loss improved the DNN fairness-performance trade-off more effectively than a fairness-regularized loss, and that GPT-4o-mini few-shot prompting outperformed the DNN baseline overall, particularly for Hispanic participants. Hispanic BA rose from $[0,24]$1 in the DNN baseline to approximately $[0,24]$2–$[0,24]$3 across the reported LLM settings, but racial disparities still persisted and neither prompt framing nor increasing the number of shots reliably removed them (Junias et al., 30 Sep 2025). Taken together, these results position DAIC-WOZ not only as a depression benchmark but also as a testbed for fairness analysis under severe subgroup imbalance.

6. Reproducibility, benchmark interpretation, and current research practice

DAIC-WOZ is often described as one of the best-known public resources for automatic depression detection, but its literature illustrates several reproducibility and interpretation problems. Some studies follow the official 107/35/47 partitioning and report development results because test labels are hidden (Burdisso et al., 2024, Watawana et al., 25 Mar 2026). Others depart from that protocol. A transcript-only classical-ML case study, for example, used an approximate 80/20 split, removed three structurally problematic interviews, and ended with 148 training and 37 test interviews, without fully documenting how this resplitting related to the official corpus partitions (Lorenzoni et al., 2024). This is an important reproducibility limitation because performance numbers obtained on custom splits are not directly comparable with official-split studies.

Feature availability and modality handling also vary by task setting. In AVEC-oriented visual work, raw video was not provided and studies relied on organizer-supplied facial behavior outputs (Dham et al., 2017), whereas other DAIC-WOZ descriptions enumerate raw audio, transcripts, and derived visual/acoustic streams as part of the corpus package (Qureshi et al., 2019). Likewise, some papers explicitly isolate participant speech or participant text using transcript timings (Dham et al., 2017, Qureshi et al., 2019), while others retain question-answer structure or interviewer prompts in the textual input (Junias et al., 30 Sep 2025). This suggests that “using DAIC-WOZ” is not a single methodological choice but a family of choices about speaker selection, segmentation, modality inclusion, and target definition.

The most robust methodological recommendations in the recent literature are therefore diagnostic rather than architectural. Studies centered on interviewer effects recommend participant-only baselines, speaker-segmented ablations, and interpretability analyses that localize evidence by speaker and normalized interview time (Burdisso et al., 2024, Watawana et al., 25 Mar 2026). Fairness work recommends explicit subgroup evaluation rather than relying on aggregate scores, particularly because DAIC-WOZ contains severe imbalance in some demographic-label intersections (Junias et al., 30 Sep 2025, Bailey et al., 2020). Across these lines of work, a high score alone is treated as insufficient evidence that a model has learned clinically meaningful depression markers from participant behavior.

DAIC-WOZ thus occupies a dual position in the literature. It remains an important benchmark for PHQ-8 regression, binary depression classification, multimodal fusion, and transcript-based modeling (Dham et al., 2017, Qureshi et al., 2019, Mao et al., 2022). Simultaneously, it has become a cautionary corpus for studying how semi-structured interview protocols, interviewer adaptation, demographic imbalance, and evaluation choices can inflate apparent performance or distort what models actually learn (Burdisso et al., 2024, Watawana et al., 25 Mar 2026, Bailey et al., 2020, Junias et al., 30 Sep 2025).