CSI–Audio Bedroom Activity Corpus: Synchronized WiFi Channel State Information and Audio for Sound Event Detection with Environmental Context

Synchronized WiFi channel state information (CSI) and audio recorded in three rooms with six participants (3 M / 3 F, 25–61 years), for sound event detection with environmental (motion) context. Companion dataset of the M.Sc. thesis Addressing False Positives in Sound Event Detection Through Environmental Context Fusion (M. Mbarki, University of Trento, 2026, supervisor Prof. Luca Turchet). Contents. 34 sessions (1 316 recorded minutes) in a fully crossed design (every participant recorded every script in every room). CSI: single ESP32-C5 → ESP32-C5 link, 100 frames/s, channel 11, HT40, MCS0, raw int8 I/Q export per frame with on-chip and host timestamps, stored as one-minute NumPy archives. Labels: timestamped span labels of body movement (walking, in_bed_activity, still, no_presence) plus markers. Metadata: session register, pseudonymous participant covariates, class mapping, protocol and room layouts. Derived: 13 519 five-second windows (18.8 h) with quality and cleaning flags and the fold assignments used in the thesis, and a 256-dimensional CED audio embedding with four audio-quality scalars per window. Protocol. Four scripts: rest and presence (A), activity coverage including weak-geometry variants (B), audio confusers — television and speech while still or absent (C), and an occupancy arc with natural transitions (D). Labels name what the body does, never the sound, so that acoustic false positives can be studied. Raw audio is not part of this record. It is archived in an open companion record (DOI 10.5281/zenodo.23459771, linked under Related works) under CC BY 4.0. See README.md for file formats, quality gates, limitations and citation.