A paired dataset of voiced and silent Korean commands with headset facial and head-motion signals

Speaking voice commands aloud raises privacy concerns in shared settings, which silent speech interfaces avoid. Without sound, these interfaces rely on articulation, such as facial movement, which can differ between voiced and silent speech, so voiced training data may not transfer. Because this difference varies between speakers, voiced–silent pairs of the same texts are needed from many speakers. Existing datasets usually provide many such pairs from one speaker or many speakers with few shared texts, and all rely on external sensors. With the built-in face tracking of a Meta Quest Pro headset, we recorded 20 Korean speakers producing the same 1,000 commands aloud and then silently, yielding 19,999 pairs. Each recording contains 63 facial movement weights and 13 head-motion channels, and voiced recordings add 19.6 hours of reference audio. Validation showed that lower-face weights reflect the command produced. The dataset supports developing silent command recognition from Quest Pro facial weights for new speakers, testing voiced recordings as additional training data and comparing facial movement between speakers and modes. Files commands.csv: the 1,000 commands with their identifier, Korean text and topic label. recording_index.csv: one row per recording, with participant code, speaking mode, command, collection day, start time, audio duration and file paths. participant_sessions.csv: recording counts, commands recorded on each collection day and voiced audio duration for each participant. P01.zip–P20.zip: one archive per participant. Each recording has a facial movement file and a head-motion file (CSV), and each voiced recording also has a reference audio file (WAV, mono, 16-bit, 44.1 kHz). The 23 files total 7.7 GB (15.9 GB uncompressed). Access The files can identify participants by voice and movement and are therefore shared under controlled access. To obtain access, request access on this page and email the signed Data Use Agreement (https://github.com/kimhaejun9568/korean-voiced-silent-commands/blob/v1.0/DATA_USE_AGREEMENT.md) to [e-mail]. Every user, including students and researchers without an institutional affiliation, signs the agreement and has their identity verified through an institutional email address or other verifiable information; students do not need a supervisor's co-signature. Access is then granted without review of the intended research, and requests are answered within 10 working days. The agreement permits research use in academic or commercial settings, prohibits attempts to identify or contact participants, and prohibits redistribution of the data and publication of audio recordings.