InterView-C shares virtual interview movements while withholding voices
InterView-C brings synchronized movement and corrected transcripts from German virtual-reality interviews into a research dataset. The single preprint describes restricted access, removal of names and withholding of raw audio. Gaze, movements and sensitive answers still carry identification risks. Tests of speech-recognition systems also expose difficulties with short answers, numbers and negation in this small sample.
Artificial Intelligence··Night
Interview movements remain linked to corrected speech
InterView-C, a dataset of virtual-reality survey interviews, links gaze, facial tracking and body movements with corrected transcripts. Patrick Schrottenbacher and colleagues describe it in a single new preprint. Two human interlocutors appear as avatars in 27 German interviews, using Quest Pro headsets. The dataset tracks both sides of the exchange rather than only the respondent. Its topics include income, employment and relationship conflicts.[1]
Raw voices stay outside the released dataset
Researchers with a non-commercial research agreement can request access. Person, place and institution names are replaced by typed placeholders, and raw audio is withheld. Participants gave informed consent for researchers to record the exchanges and share research material with identities removed. The authors nevertheless identify remaining re-identification risks from movements, gaze and sensitive answers. With audio unavailable, outside researchers cannot test new speech-recognition systems on the same recordings.[1]
Short answers and negation expose recognition errors
Speech-recognition experiments covered nine systems. Errors were particularly frequent when recognizing brief responses, spoken numbers and words expressing negation. Human annotators corrected 54 transcripts while retaining pauses, repetitions and false starts. The sample contains nineteen male and eight female interviewees, all highly educated, alongside three interviewers from the project team. Interviewer avatars were not randomized. The authors limit their account of interview behavior to exploratory findings from this small group, whose participants are unevenly distributed.[1]
