InterView-C contains synchronized gaze, face, head, body and hand tracking of interview participants together with their transcribed speech and their answers to sensitive survey questions. Access is granted under a data use agreement: the data may be used for non-commercial research only, and any attempt to re-identify participants, alone or by linking with other data, is prohibited. Requests are reviewed manually.

InterView-C

A German multimodal corpus of 27 survey interviews conducted entirely in virtual reality. Interviewer and interviewee meet as avatars in a shared virtual room; the Meta Quest Pro headsets of both record gaze, face, head, body, hand and finger tracking alongside their speech.

The corpus has three parts:

VR interaction data — synchronized tracking of both speakers, word-timed automatic transcripts, the time windows in which each questionnaire item was read out, and the questionnaire data.
Reference transcripts — a manually post-edited verbatim transcript of all 54 recordings (one per speaker and interview), forced-aligned to the same clock.
Negation annotations — negation cue and scope labels for 1,422 transcribed sentences (1,398 doubly annotated), with train/validation/test splits.

The raw audio is not part of the release.

Loading

from datasets import load_dataset

words = load_dataset("AnonymousKiwi/InterView-C", "reference_word", split="train")
gaze  = load_dataset("AnonymousKiwi/InterView-C", "eye", split="train")

Every table is a single Parquet file, so pandas.read_parquet / DuckDB / Polars on the files under data/ work as well.

Time base

All tables share one session clock: timeMs is integer milliseconds since the Unix epoch, time is the same instant as a UTC timestamp. Each headset's clock was mapped onto a shared clock with one constant offset per session; sub-second timing follows the audio sample clock. Tracking frames, words and item windows can therefore be joined on time directly, e.g. the interviewee's gaze during an answer:

import pandas as pd
w   = pd.read_parquet("data/transcripts/ReferenceWord.parquet")
eye = pd.read_parquet("data/tracking/Eye.parquet")
ans = w[(w.experiment == 12) & (w.role == "Interviewee") & (w.line == 40)]
g   = eye[(eye.experiment == 12) & (eye.role == "Interviewee")
          & eye.timeMs.between(ans.timeMs.min(), ans.timeMs.max())]

Every row carries experiment (interview id) and role (Interviewer / Interviewee). playerId identifies a headset session: a participant who reconnected has several sessions within one role (see Player.playerIds).

Tables

Config	File	Rows	Content
`experiment`	`session/Experiment.parquet`	27	interview start/end, `interviewerId` (anonymous code 3/4/5)
`player`	`session/Player.parquet`	54	one row per speaker and interview; `playerIds` lists every headset session
`eye`	`tracking/Eye.parquet`	1,350,572	binocular gaze poses, validity, confidence
`head`	`tracking/Head.parquet`	1,350,572	6-DoF head pose
`body`	`tracking/Body.parquet`	1,350,572	6-DoF rig (body root) pose
`facial`	`tracking/Facial.parquet`	1,350,572	63 face-expression weights
`left_hand` / `right_hand`	`tracking/{Left,Right}Hand.parquet`	1,350,572 each	6-DoF hand/controller anchor pose
`left_finger` / `right_finger`	`tracking/{Left,Right}Finger.parquet`	1,350,572 each	hand-tracking skeleton, root/pointer pose, pinch, confidence
`word`	`transcripts/Word.parquet`	53,475	automatic word tokens (CrisperWhisper), time-aligned
`reference_word`	`transcripts/ReferenceWord.parquet`	55,764	post-edited reference word tokens, forced-aligned
`interview_item`	`instrument/InterviewItem.parquet`	56	scripted question stems and battery statements
`question_window`	`instrument/QuestionWindow.parquet`	1,147	when each item was read out, located by aligning the script to the interviewer transcript
`survey_item`	`questionnaire/SurveyItem.parquet`	131	items and value labels of the three questionnaires
`survey_response`	`questionnaire/SurveyResponse.parquet`	27	one response set per interview, avatars chosen
`survey_answer`	`questionnaire/SurveyAnswer.parquet`	3,537	answers with typed missingness
`negation`	`negation/sentences.parquet`	1,422	tokens, consolidated cues/scopes, split
`negation_annotations`	`negation/annotations.parquet`	2,820	the individual annotations (A1–A3)

Plain-text versions of both transcript layers (one file per recording) are in transcripts/text/{reference,automatic}/<experiment>_<Role>.txt.

Tracking

Logged at a median of 18.4 Hz; counter is the frame index within a headset session. Coordinate frames differ by stream:

Body, Head, LeftHand, RightHand: Unity transforms, world space (position x/y/z in metres, rotation quaternion w/x/y/z). Hand is the hand/controller anchor.
Eye, Facial, LeftFinger, RightFinger: raw OVRPlugin structs in the native right-handed tracking space. To get Unity's convention, flip position (x, y, -z) and rotation (-x, -y, z, w). deviceTime (Eye, Facial) is the headset's own sample time, not the session clock.
Facial.expressionWeights holds the 63 weights in OVRPlugin.FaceExpression order; status.IsValid flags valid frames.
Eye.eyeGazes is [left, right], each with Confidence, IsValid and Pose.
*Finger has one row per frame. status is the OVRPlugin.HandStatus bit field: a hand is tracked when status & 1 (HandTracked), which holds for 38.2 % (left, 515,995) and 40.3 % (right, 544,280) of frames — the rest of the time the participants held controllers. Use that bit, not the presence of boneRotations: frames with only InputStateValid set still carry bones. handConfidence is the raw float bit pattern (1065353216 = 1.0, high). boneRotations are the 26 OpenXR hand joints (XR_HAND_JOINT_* order) as absolute rotations in tracking space, not parent-relative.

Transcripts

Word is the automatic CrisperWhisper transcript; ReferenceWord is its manual verbatim post-edit (filled pauses, repetitions and false starts kept; numbers up to thirteen spelled out). Both are word tables with index (running position per recording), text, timeMs/time, endTime, duration.
ReferenceWord.line groups words into the annotators' transcript lines; alignScore is the mean character probability of the CTC forced alignment (German wav2vec 2.0). 0.3 % of reference words could not be aligned and have no time. Markers for unintelligible speech ([unknown]) have no audio to align to and appear only in the plain-text reference files.

Questionnaire

SurveyItem.source is faces (55 interview items, incl. three recording-consent prompts), v1309 (50 items, interviewee's experience questionnaire) or v1261 (26 items, interviewer's protocol). SurveyAnswer has valueNum + valueLabel for coded answers, value for free text and missing for missing answers: filtered (not asked, filter skip), dont_know, refused, no_answer, not_collected (questionnaire part missing). The open main-questionnaire items (occupation, dream occupation, …) are coded only as answered/not answered; their content is in the transcripts.

Avatars in SurveyResponse are coded A–D: A blue hair, B striped pullover, C sunglasses, D headscarf.

Negation

Following the cue/scope formulation of *SEM 2012: negation/sentences.parquet holds per sentence its tokens, split (train/val/test, by whole interview: 21/3/3), speaker role, and the consolidated groups (one per negation instance: cueIndices, scopeIndices, how the scope was decided). Cues are the union of both annotators; disputed scope tokens were resolved with a Dawid–Skene model fitted on the training split (negation/aggregation_report.json). cueUnanimousIndices is the stricter cue alternative. CoNLL files for training are under negation/conll/<split>/: reference.{cue,scope}.conll, baseline_cue_unanimous.cue.conll, baseline_scope_{union,intersection}.scope.conll; the .json next to each maps CoNLL blocks to sentenceId.

De-identification

No audio, no images, no names of participants.
In all transcript layers, names of persons, places and institutions are replaced by typed placeholders [PERSON], [LOCATION], [ORGANIZATION] — one placeholder per original token, so word counts and timing are unchanged.
Interviewers are identified only by an anonymous code, annotators as A1–A3.
Head and hand motion can re-identify VR users given reference recordings of the same person, which is why access is restricted (below).

Access and data use agreement

The dataset is available on request. By requesting access you agree to the following terms:

The data is used for non-commercial research only.
You do not attempt to re-identify any participant, neither from the data alone nor by linking it with other data, and you report any accidental re-identification to the dataset maintainers without disclosing it further.
You do not redistribute the data or derived data that would allow reconstructing it; share access by pointing others to this page.
You keep the data secure and delete it once your research purpose ends.
Publications using the data cite the InterView-C paper.

The raw audio of both speakers is withheld and is not available on request.

Downloads last month: 3

Total file size:

1.18 GB